Market Accessibility, Corporate Bond ETFs,

and Liquidity∗

Craig W. Holden

Kelley School of Business

Indiana University

Jayoung Nam

Cox School of Business

Southern Methodist University

March 18, 2019

∗We thank Robert Battalio, Azi Ben-Rephael, Utpal Bhattarcharya, Matthew T. Billett, Dion Bon-gaerts, Brittany Cole, Caitlin Dannhauser, Richard Evans, Thierry Foucault, Vincent Glode, Eitan Gold-man, Ryan D. Israelsen, Sreenivas Kamma, Vincent van Kervel, Pete Kyle, Mina Lee, Katya Malinova,Rabih Moussawi, Christine Parlour, Veronika K. Pool, Uday Rajan, Jonathan Reuter, Jan Schneemeier,Batchimeg Shambalaibat, Philip Strahan, Noah Stoffman, Charles Trzcinka, S. “Vish” Viswanathan, ZhenyuWang, Wenyu Wang, Avi Wohl, Jun Yang, Wei Yang, Scott Yonker and seminar participants at In-diana University, the FIRS 2017 meeting, Eighth Erasmus Liquidity Conference, The Financial TheoryGroup Summer School, the Northern Finance Association, the Financial Management Association Meet-ings, and the 14th WFA-CFAR Corporate Finance Conference. Contact address: Jayoung Nam, SMU CoxSchool of Business, 6214 Bishop Blvd, Dallas, TX 75275; E-mail address: jayoungn@smu.edu; Webpage:https://sites.google.com/view/jnam/home

1

Market Accessibility, Corporate Bond ETFs, and Liquidity

Abstract

We find that market accessibility ex ante plays an important role in how the underlying

assets’ liquidity changes when a basket security is introduced. First, using a multi-market

version of the Kyle model, we show that the less (more) accessible the underlying market

is, the more its liquidity improves (deteriorates) when basket trading becomes available.

Second, we empirically test these predictions using corporate bonds before and after the

introduction of ETFs. Consistent with the model, liquidity improvement is larger for highly

arbitraged, low-volume, high-yield, and long-term bonds and for 144A bonds to which retail

investor access is prohibited by law.

2

1 Introduction

Prior literature on the introduction of basket securities shows that the liquidity of the un-

derlying assets deteriorates because uninformed investors migrate to the basket market in

order to reduce the cost of trading against informed traders. In this paper, we show that if

investor participation in the underlying market is limited, the liquidity of the underlying se-

curities can, in fact, improve after basket securities begin trading. To do so, we first develop

a theoretical model and find that the level of market participation ex ante is an important

determinant of how the underlying assets’ liquidity changes ex post. Second, we empirically

test the predictions of the model by investigating the liquidity of corporate bonds before

and after the introduction of corporate bond exchange traded funds (ETFs) using bond

transaction data from TRACE.

Theories of basket securities trading typically rely on the Kyle model (Kyle, 1985), in

which prices partially reflect informed investors’ information. In the model, liquidity trading

demand is independent of the private information about individual securities. Therefore,

liquidity investors are concerned with their expected losses from trading due to adverse

selection. Since basket securities diversify away the idiosyncratic risks of the underlying

securities, these liquidity investors can reduce their expected losses by trading in basket

securities. Hence discretionary liquidity investors migrate to the basket market and conse-

quently, liquidity in the underlying asset market declines. (Subrahmanyam, 1991; Gorton

and Pennacchi, 1993; Jegadeesh and Subrahmanyam, 1993; Hamm, 2014; Israeli, M. C. Lee,

and Sridharan, 2017)

Largely overlooked is the fact that the basket market and the underlying market may

not be equally accessible for liquidity investors. A large body of literature has shown that

investors’ market participation can be limited due to unfamiliarity (Merton, 1987; Huberman,

2001), home bias (Kang and Stulz, 1997; Pool, Stoffman, and Yonker, 2012), regulatory

constraints (Ellul, Jotikasthira, and Lundblad, 2011), lack of information, high transaction

3

costs, minimum investment requirements, or difficulty in liability matching. This implies

that for liquidity investors, diversification can be quite costly and trading barriers can be

high. As a specific example, for retail investors, the transaction cost associated with trading

an individual junk bond averages 110 basis points (Figure 1). Furthermore, the minimum

trading size for individual bonds is $100,000 at a major brokerage firm. In sharp contrast,

with the introduction of ETFs, investors can trade an entire corporate bond portfolio at a

cost of 1 basis point with an investment as small as $1,000. Additionally, an example of a

fully inaccessible market is 144A bonds, in which trading by retail investors is not legally

permitted.

To derive theoretical predictions, we develop a multi-market, adverse selection model

based on the adverse selection models of Kyle (1985) and Subrahmanyam (1991). In our

model different markets have different levels of accessibility (defined more precisely below).

We show that the introduction of basket securities in a market with limited investor partic-

ipation leads to an improvement in the liquidity of the underlying securities. An important

mechanism that links limited market participation to liquidity is the entrance of new unin-

formed investors into the basket market. This happens because they can achieve a better-

diversified portfolio at a low cost through basket securities. The high liquidity of the basket

securities spills over to the underlying market by arbitrage trading, which we incorporate by

integrating the arbitrage trading model of Holden (1995). This liquidity spillover is greater if

the influx of these new uninformed investors outweighs the departure of uninformed investors

from the underlying market, which is likely to occur when the underlying market ex ante is

less accessible. Conversely, if the underlying market ex ante is highly accessible, then the

departure of uninformed investors from the underlying market exceeds the entrance of new

investors and the liquidity of the underlying securities deteriorates.

Testing whether trading basket securities in a market with limited participation improves

or hurts liquidity of the underlying securities requires that a proper financial instrument be

available. Existing theoretical and empirical studies have focused on the equity market and

4

unanimously find only a one-directional effect: the deterioration of liquidity in the underlying

markets upon the introduction of a basket security. These results are consistent with one of

the predictions of our model since the underlying equity market is already highly accessible.

Yet few studies have been conducted outside the equity market, for example in the bond

market, due to limited data availability of the instruments. One possibility would be open-

end bond mutual funds, which, like bond ETFs, allow investors to achieve a diversified

portfolio at a smaller cost than investing in a portfolio of individual bonds,1 and have long

been available. Although there are no price differences between bond mutual funds and the

underlying assets,to the extent that arbitrageurs primarily utilize trading volume differences

between two correlated markets, the liquidity spillover mechanism would still be present

in bond mutual funds. Therefore, a first environment to test our theoretical predictions

would be around 1986 when Vanguard introduced the first bond index mutual funds.2 Un-

fortunately, the transaction data for corporate bonds are unavailable. A second possibility

would be around 2004 when the TRACE dataset became publicly available. However, newly

launched mutual funds nearly 22 years after the introduction of initial bond mutual funds

may not be an ideal environment in which to assess the effect of market accessibility.

In this paper, we test the predictions of our model using TRACE intraday trading

data for corporate bonds and ETFs since 2007.3 Using the indicator-based regression

methodology (Huang and Stoll, 1997; Bessembinder, Maxwell, and Venkataraman, 2006)

in a difference-in-difference (DD) framework, we compare the changes of the effective spread

of corporate bonds before and after they are included in a newly launched ETF for the first

time. Our control sample consists of corporate bonds that are not included in ETFs but are

similar along other dimensions such as credit rating, time to maturity, issue size, issuer, etc.

1Passively-managed high yield bond mutual funds are non-existent. Actively-managed high yield bondmutual funds cost 138 basis points annually. The annual costs for investment grade bond mutual funds arelower: 88 basis points if actively managed, 33 basis points if passively managed (Cici and Gibson (2012);Chen, Ferson, and Peters (2010)).

2Vanguard total bond market index mutual fund3While equity ETFs started trading in 1993, corporate bond ETFs became available only after 2002 and

have only become popular since 2007.

5

We find that non-investment grade ETF bonds experience improved liquidity (i.e., a

reduction in transaction costs) of 10%, while investment grade ETF bonds experience an

insignificant change in trading costs. ETF bonds with low volume see improved liquidity

once the ETF starts trading, while those with large volume see decreased liquidity. ETF

bonds whose remaining life is greater than 2 years see improved liquidity. As evidence for

the underlying economic mechanism, we estimate the degree of arbitrage between the ETF

market and the underlying bond markets using the creation and redemption activities of

ETFs. We find that bond liquidity improves only when the arbitrage activities are high.

An especially interesting cases is that of 144A bonds, which are privately placed corporate

bonds and are traded among a limited number of qualified investors (i.e., large institutional

investors), such as insurance companies (Carey, Prowse, Rea, and Udell, 1993). Retail

investors are prohibited by law from trading 144A bonds. However, they can gain exposure

to 144A bonds for the first time once bond ETFs start holding them. We find that the

underlying market in 144A bonds experiences a 20% improvement in liquidity once bond

ETFs start holding 144A bonds. This further supports our key theoretical prediction. That

is, if market accessibility is even more limited, then uninformed demand for ETFs can have

an even more positive liquidity spillover to the underlying bonds.

To further support the causality of our results, we also use a quasi-natural experiment

that happened to one of the major high yield bond ETFs, namely, the iBoxx high yield $

corporate bond ETF (HYG). In June 2009, the index composition rules were changed and

were announced only 8 days before the effective date. As a consequence, the number of newly

added bonds dramatically increased to 248 versus only 16 bonds were added previously from

January to May of 2009. Consistent with our predictions, we find an improvement in the

liquidity of those newly added bonds of 3.5%.

Our findings contribute to our understanding of the liquidity of assets by providing new

evidence about the role of market participation, the channel through which it operates, and

the magnitude of its effects on the liquidity of the underlying assets. Both our theoretical

6

results and our empirical tests using bond ETFs are unexpected, given the past evidence

that trading basket securities causes the liquidity of the underlying assets to deteriorate.

However, as previously noted, the existing literature focuses only on the U.S. equity market,

which is a highly accessible market.

Extant literature on the liquidity of the bond market has sidestepped the role of investor

market participation, and has focused on the supply side of the market. While empirical

studies find that the liquidity of corporate bonds can be affected by innovative market mech-

anism design (Biais and Green, 2019; Abudy and Wohl, 2017; Hendershott and Madhavan,

2015), by regulation of post-trade transparency (Schultz, 2001; Bessembinder et al., 2006;

Edwards, Harris, and Piwowar, 2007), and by dealers’ capital commitment (Bessembinder,

Jacobsen, Maxwell, and Venkataraman, 2017; Choi and Huh, 2017), market accessibility is

assumed to be unaffected – perhaps because market participation is hard to change due to

large trading barriers to entry. Our findings suggest that when the underlying market ac-

cessibility is more limited, the liquidity spillover benefits of ETF trading are greater. Thus,

one fundamental driver of corporate bond liquidity appears to be market accessibility.

Our paper also contributes to the growing literature on ETFs. We investigate whether

the introduction of ETFs has any impact on the liquidity of the underlying assets and find

that the effects can be predicted based on long term market accessibility ex ante. Con-

currently, there is active interest in various other impacts of ETFs on market quality: on

market fragility (Bhattacharya and O’Hara, 2016), on the fragility of liquidity spillovers (Pan

and Zeng, 2017), on market volatility (Ben-David, Franzoni, and Moussawi, 2018), on co-

movements of asset returns (Da and Shive, 2018), and on the valuation of the underlying

assets (Dannhauser, 2017). The important aspect of our paper is that market quality can

change in opposite directions depending on long term market accessibility ex ante.

The two papers closest to our study are Dannhauser (2017) and Ye (2019). Dannhauser

(2017) examines how ETFs affect the value of the underlying bonds. Using the same natural

experiment, she shows that corporate bond ETFs improve bond valuations and argues that

7

this is due to the influx of institutional investors into the ETF market. As an auxiliary result,

she also finds that bond market liquidity is not significantly affected overall. A potential

source of the discrepancy between the Dannhauser results and those in this paper stems

from differences in methodology. Our study employs the event study method six months

before and after the introduction of ETFs using the bonds added to the ETF portfolio for

the first time, while her study utilizes a long panel regression over four year periods with

ETF ownership.

In a contemporaneous working paper, Ye (2019) empirically examines the liquidity of

corporate bonds when they are included or excluded from corporate bond ETFs. Using a

liquidity measure which is different from that in this paper, he finds that corporate bond

liquidity improves when a bond is included in an ETF and deteriorates when it is excluded.

His results corroborate our findings and suggest that the argument that the introduction of

ETFs enhances bond liquidity is robust to different measures of liquidity and is symmetric

to exclusions. However, the theoretical contribution of our paper explains why underlying

liquidity is expected to respond differently to the introduction of ETF trading in the stock

and bond markets.

The rest of the paper is organized as follows. In section 2, we develop our theoretical

model and derive empirical predictions. In section 3, we discuss the data and the indicator-

based regression methodology. In section 4, we discuss the empirical results. In section 5,

we conclude.

2 The Model

The first step of our analysis is to model trading of liquid assets as a means of ensuring

uninformed investors’ ability to maintain a diversified portfolio efficiently. Our basic model

is a simple representation of a dynamic problem in which all the investors (uninformed,

informed, arbitrageurs, market makers) optimize their trading decisions and in which new

8

liquid assets increase accessibility of the market. Depending on market accessibility ex ante,

the introduction of new liquid assets may have positive, negative, or neutral impacts on the

liquidity of the assets in place.

2.1 The setup

We consider a security market with two assets and use i ∈ {1, 2} to index the assets. There

are two times. Trading takes place at time 0 and the assets are liquidated at time 1. Let

vi be the time 1 liquidation value, which we assume is normally distributed, vi ∼ N(µi, σ2i ).

For simplicity, we assume that the two assets are uncorrelated, Corr(v1, v2) = 0.4

As in the standard Kyle model, there are (i) informed investors, (ii) uninformed investors,

and (iii) market makers. There are two types of uninformed investors, discretionary or non-

discretionary, as in the Subrahmanyam model.

There are ki risk neutral, informed investors who specialize in the ith security market,

and who have private information about the idiosyncratic risk εi ≡ vi − µi of security i.

They have no interest in speculating in the markets for other securities because ρ = 0.

Informed investor j (j = 1, · · · , ki) in the ith security market optimizes her expected profit,

E [xij (Pi − vi) |εi], from trading on her private information by choosing her optimal demand,

xij, where the price Pi of the security will be determined by the market makers.

There are non-discretionary uninformed investors in each of the markets, who are ob-

ligated to trade a particular security and are not allowed to trade other assets. Their de-

mand for security i is denoted by ui. We assume that this demand is normally distributed,

ui ∼ N(0, σ2ui

).

There are Q risk averse, discretionary liquidity investors who may invest in either of

the assets. They have to trade the assets due to random arrival of liquidity needs. Let lq

denote the demand of investor q (q = 1, · · · , Q), which we assume is normally distributed

lq ∼ N(µl, σ2l ). In order to keep their portfolios diversified, they would prefer to allocate their

4All of the results continue to hold for ρ 6= 0.

9

trading need across the assets with weights αi, to be determined. However, high transaction

costs and high minimum trading requirements may prevent this optimal allocation.

There are multiple risk neutral, market makers in each market. The market makers

observe the demand from all investors – informed, non-discretionary, and discretionary –

and set the price (Pi) of asset i. Competition among the market makers implies that they

set the prices of the securities so as to earn zero excess profit. Following the linear pricing

rule as in the Kyle model, the price of each security i is set by the market makers as

Pi = µi + λiωi = E [vi|ωi]. As in the Subrahmanyam model, the aggregate order flow ωi

is augmented with additional trading demand αilq from discretionary uninformed investors.

The total informed trader demand is xi =∑ki

j=1 xij. The slope of price function λi scales

the price impact of the aggregate order flow submitted by all agents and is determined

endogenously within the model.

We assume discretionary investors have a negative exponential utility function, − 1γ

exp(−γW ),

over their time 1 wealth W , where γ > 0 is the risk aversion parameter. In order to maximize

their diversification benefit, the discretionary investors may choose to split their demand lq

between the assets. But they are concerned about losses due to trading against informed

investors. The expected loss from trading security i will be E [(Pi − vi)αilq] = λiα2i Var(lq).

The uninformed investors wish to maximize their expected utility,

maxα1,α2

E[−1

γexp (−γW )

]= max

α1,α2

{−1

γexp

[−γE(W ) +

γ2

2Var(W )

]}, (1)

W =

(α1W0

µ1

+ α1lq

)v1 +

(α2W0

µ2

+ α2lq

)v2 − (P1 − v1)α1lq − (P2 − v2)α2lq, (2)

E [|α1lqP1|] ≥ B1, for lq 6= 0, and E [|α2lqP2|] ≥ B2, for lq 6= 0, (3)

where 0 ≤ α1, α2 ≤ 1. The objective function (1) is feasible only if the optimal trading

allocation decisions are feasible. We assume in (3) that the expected trading in each security

market must exceed a minimum trading barrier. The minimum trading barriers Bi > 0 are

10

exogenously provided.5 Positive lq indicates buying. Negative lq implies selling.6 If lq = 0,

then no trading occurs, in which case the strategic decision is reduced to the traditional

setting for long term investment.

Definition 1. The equilibrium of the model consists of:

(i) a trading allocation strategy (α1, α2) by discretionary uninformed investors that maintains

the diversification benefits, given the price functions and the trading strategies of informed

investors and non-discretionary uninformed investors,

(ii) a trading strategy (x1, x2) by the informed investors to maximize their expected trading

profits, given the price functions and the trading strategies of the other agents, and

(iii) the price function (P1, P2) set by market makers with zero excess profits.

In order to analyze the effect of the introduction of ETFs on the equilibrium, the baseline

model above is modified as follows. The value of the ETF at t = 1 depends on the values

of the underlying assets, vE = w1v1 + w2v2, where wi is the weight if security i in the

ETF and the weights sum to one. The informed investors in the market i can use their

private information about vi to draw inferences about the value of ETFs, although the

signal is noisy. They trade the ETF to maximize their expected profits. We assume non-

discretionary uninformed investors have an exogenous trading demand uE for the ETF that

follows a normal distribution N(0, σ2uE

). The discretionary uninformed investors reconsider

their strategic decisions on the optimum portfolio trading allocation across securities 1, 2,

and the ETF, i ∈ {1, 2, E}. The trading barrier for ETF is strictly smaller than those for

the individual securities, BE < B1 and BE < B2. For simplicity, we assume the required

minimum trading for the ETF is nonexistent: BE = 0. ETF market makers set the price

with zero excess profits due to competition.

5In reality, the trading barriers Bi take both monetary and non-monetary forms and their magnitudesvary. Trading individual stocks costs little so that Bi is effectively zero. In contrast, the minimum tradingsize for regular corporate bonds is $100,000, while 144A bonds trading are not legally permitted to retailinvestors so that Bi = ∞. Investors are less familiar with high yield bonds than investment grade bondsthat Bi for high yield bonds is higher than that for investment grade bonds.

6The likelihood of selling can be made small by assuming µl > 0, but the results of our paper remain thesame.

11

In addition, arbitrageurs exploit price differences, PE − w1P1 − w2P2, between the un-

derlying securities and the ETF. This is possible because ETFs have their own supply and

demand in the market, independent of those for the underlying securities. The price PE for

the ETF can diverge from the prices of the underlying assets. We assume arbitrageurs’ trad-

ing demand θ is proportional to the trading volume difference of the uninformed investors

uE +αelq−α1(u1 + lq)−α2(u2 + lq). They trade to maximize their arbitrage trading profits.

Finally, the new equilibrium of the model incorporates the optimal trading activities θ∗

by arbitrageurs.

2.2 Analysis

2.2.1 Unconstrained solution. The market is accessible when the uninformed investors

are able to trade at their best levels before and after the ETFs are introduced. In the pre-

ETF period, optimal trading allocations are found to be

α?i =1

γ

(W0 −

σ2l

Ai+ γ

σ4l

Ai

)[W 2

0 σ2i + µ2

iσ2l (σ

2i + 1)− σ2

l µ2iσ

2i

(k2i + ki + 1

k2i + 2ki + 1

)]−1

,

where Ai =[

ki(ki+1)2

µ2i σ2i

σ2u+σ

2l

]−1/2

. See Appendix A for the proof. The assumption of accessibil-

ity means that the investors’ trading strategy α?i satisfies the trading constraints (3). More

explicitly, α?i satisfies |α?i lqPi| ≥ Bi for i = 1, 2. The exogenous parameters that determine

the market accessibility are the investors’ liquidity trading demand lq and a minimum trad-

ing requirement Bi. If investor q’s liquidity demand is large relative to the required trading

minimum, then that investor is not constrained and can access the market. A market is

accessible to all investors if the trading minimum Bi is small enough. One of the important

endogenous parameters is the impact λi of trading by all market participants. Under the

assumption that market i is accessible to all participants,

λi =

√ki

(ki + 1)2µ2iσ

2i

σ2ui

+ α?2i σ2l

,

12

which is smaller than it would be if the market were partially accessible. This implies that

the market is more liquid.

In the post-ETF period, the optimal trading allocation is found to be α̃?i = α?i − α̃?ewi.

This post-ETF trading allocation α̃?i in the accessible markets is smaller than the pre-ETF

trading strategy α?i , i.e. α̃?i < α?i . (See Appendix A for proofs of these statements.) The

ETFs are basket securities where idiosyncratic risks have been diversified away. The cost of

and probability of trading against an informed investor, also known as adverse selection, are

lower in the ETF market. As a consequence, discretionary uninformed investors prefer to

trade ETFs. The endogenous liquidity parameter λ̃i is found as

λ̃i =

√ki

(ki + 1)2µ2iσ

2i

σ2ui

+ α̃?iσ2l + σ2

θ + 2 Cov(θ, ui).

Under the condition that σ2θ + 2 Cov(θ, ui) < (α?i − α̃?i )σ2

l , the post-ETF liquidity declines

as the trading leads to a bigger price impact, i.e. λ̃i > λi. For our purposes, the main

implication of this condition is that if the market is highly accessible, the trading volume σ2l

by discretionary uninformed investors is large and this condition is more likely to hold.

2.2.2 Constrained solution. The market is inaccessible when the discretionary unin-

formed investors are not able to achieve their best trading strategy, α? for i = 1, 2 because

of market accessibility frictions. When the market is inaccessible, the uninformed investors’

optimal trading strategy above may not be implemented. More explicitly, the condition for

the market to be inaccessible is that for some realizations lq, the trading strategy α? above

does not satisfy the condition

α?i =BiAi − (xi + ui)lq

µiAilq + l2q.

We assume that when the optimal trading strategy is infeasible, an uninformed investor q

trades the risk free security. As a consequence, some uninformed investors hold an under-

13

diversified portfolio in the pre-ETF period. In the extreme case that the optimal strategy is

infeasible for all discretionary ‘uninformed investors, the market liquidity in this market is

related to the non-discretionary uninformed investors only,

λi =

√ki

(ki + 1)2µ2iσ

2i

σ2ui

.

We assume the minimum trade constraint Be for the ETF market is zero for simplicity.

The introduction of ETFs allows these uninformed investors to trade the ETFs at their

best level α?e (See Appendix A for the proof). This influx α?elq of the uninformed investors

further increases the ETF liquidity, and increases the arbitrage trading volume σ2θ . The ETF

liquidity spills over to the underlying market through arbitrage. The post-ETF liquidity λ̃

in the underlying market, which is defined as

λ̃i =

√ki

(ki + 1)2µ2iσ

2i

σ2ui

+ σ2θ + 2 Cov(θ, ui)

,

is smaller than its counterpart λi in the pre-ETF market. Hence market liquidity in the

inaccessible market improves after the introduction of ETFs. 7

2.3 Implications

The proposition below states the main implications of our model. A graphical representation

of the proposition is shown in Figure 3.

Proposition 1. Let the market liquidity before and after the introduction of ETFs be λi and

λ̃i, respectively.

1. If the market is sufficiently accessible ex ante, then the market liquidity decreases after

7While we assume the investors in inaccessible market reside in the risk free asset, some of the investorsmay have invested in mutual funds. If these investors migrate to ETFs from mutual funds, then potentiallythe liquidity of the bonds owned by mutual funds, exclusive of those owned by ETFs, may decline. (Changand Yu (2010)) More importantly, our main theoretical predictions on the liquidity of the underlying bondsowned by ETF, exclusive of mutual funds, remain the same.

14

the introduction of ETFs, λ̃i > λi.

2. If the market is inaccessible ex ante, then the market liquidity in the post-ETF periods

increases, λ̃i < λi.

3. As the accessibilities differ more between the basket and underlying markets, the decre-

ment (or the increment) of the market liquidity is larger.

In empirical terms, the proposition above implies that the introduction of ETFs can

have differential impacts on the stock and corporate bond markets. The stock market is

as highly accessible as the ETF market. When stock ETFs are introduced, the liquidity of

the underlying stocks should decrease. This prediction is consistent with the prior litera-

ture (Subrahmanyam (1991)) and has been tested multiple times with stock index futures

and with stock ETFs. In contrast, the liquidity of the bond market should increase when

bond ETFs begin trading. Bond ETF liquidity spills over to the underlying bond market

through arbitrage. Our theoretical predictions are driven by a comparison of accessibility

between the ETF market and the underlying securities markets. Our empirical tests that

follow are focused on less accessible bond market.

A special case of Proposition 1 can be applied to mutual funds. Arbitrageurs may pri-

marily be interested in exploiting trading volume differences between the basket market and

the underlying securities market. As long as the arbitrage trading demand θ is present, the

predictions in Proposition 1 still hold.

3 Research Design

3.1 Identification strategy

To test empirically the impacts of ETF trading on liquidity of constituent securities with

limited accessibility, We identify a source of plausibly exogenous variation in changes in

market accessibility. Most of our analysis focuses on inclusions of corporate bonds in ETF

portfolio as the result of new ETFs inceptions.

15

Identification requires that such corporate bond inclusions correlate with the introduction

of ETFs but not otherwise correlate with the bonds’ expected liquidity. A concern is that

inclusion of a security in a tradable portfolio can be viewed implicitly as a revelation of its

expected liquidity. Managers of passively managed funds would prefer to include more liquid

securities in order to minimize any tracking errors since large observed tracking errors could

deter future investors. Similarly, the managers may drop a security to which tracking errors

are attributed.

We avoid these biases by focusing on introductions of new ETFs, rather than selective

inclusions of individual bonds in existing ETFs. The period since 2002 has seen a gradual

increase in the number of fixed income ETFs. We argue that this growth has been fueled by

unexpected economic events. According to the CEO of Barclays Global Investors (BGI), the

first four corporate bond ETFs8 - LQD, TLT, IEF, SHY - were launched as an experiment.

The hope was that investors would be attracted to financial instruments featuring easy and

inexpensive access to diversified portfolios with simple governance. BGI, being the first

comer, felt that they had a comparative advantage in costs and the offering of new market

access. The bursting of the internet bubble and the Enron accounting scandal in early 2000–

2001 revealed that active fund managers are largely concentrated only in a small number of

securities. This increased investor appetite for diversified portfolios with simple governance,

and hence to survival of the very first bond ETFs. Starting in 2008, the unexpected col-

lapse of Lehman Brothers raised concern about counterparty risk with broker-dealers, upon

which interest in ETFs with more illiquid, instead of liquid, bonds developed. Finally, a

proposed 2016 FINRA fiduciary rule change affecting financial advisors has led to increased

recommendations for ETFs for the benefit of low cost diversification9.

Even so, one may worry that advisory boards for passive ETFs use a random sampling

indexing strategy, which invests in a representative sample of bonds in the underlying index,

8iShares IBoxx $ Investment grade corporate bond ETF (LQD), iShares 20+ Year Treasury BondETF (TLT), iShares 7-10 Year Treasury Bond ETF (IEF), iShares 1-3 Year Treasury Bond ETF (SHY)

9Cite the article with the chart

16

allowing some selection effects. We argue that this concern is largely muted for these new

passively managed ETFs as the composition of the underlying index is predetermined inde-

pendently from the introduction of ETFs. Although the ETF managers wish to minimize

the tracking errors, the lack of their track record in the past makes it difficult to control

for the expected liquidity of the underlying bonds. In addition, according to SEC filings

for passively managed ETFs, the selection is determined by a broad range of criteria. The

liquidity needs are addressed by allowing the funds to invest in derivatives securities or other

liquid assets defined by the advisory board. For instance, the SEC filing for one passively

managed ETF – iShares high yield bond ETF (HYG) – explains the guidelines for inclusion

of a security in the ETF portfolio as follows.10

The selected bonds have, in the aggregate, market capitalization, industry weighting,

return variability, earnings valuation, duration, maturity, credit rating, yield, and

liquidity measures similar to those of the underlying index. Furthermore, to help

ETFs track its underlying index, ETFs invest at least 80% of its assets in compo-

nent securities of their underlying index, and may at times invest up to 20% of its

assets in futures, swap contracts, cash, cash equivalents that are not included in the

underlying index.

One exception is when the benchmark index for an ETF changes its composition. In

section 4.2.2, we investigate the inclusion of corporate bonds in an existing ETF portfolio,

which was caused by the changes of the index composition rule. Since the market segment

covered by this ETF has been accessible since its inception, the empirical results depend on

how much the newly added bonds are related to the existing bonds.

The use of the introduction of new bond ETFs to identify exogenous changes of market

accessibility is new to the literature. The identification strategy behind the addition in ETF

portfolios is similar in spirit to those of ? or Dannhauser (2017). However, we believe that

our identification more cleanly identifies more true effect of ETF inclusion. Dannhauser

10https://www.sec.gov/Archives/edgar/data/1100663/000119312507007582/d485apos.htm

17

includes the bonds added to the ETFs, but, unlike me, she did not condition on those newly

launched. Ye investigates both the inclusion and exclusion effects of ETFs. But it can

capture some endogenous effects as the existing ETFs have observed their tracking errors

and can strategically decide to include more liquid bonds or to exclude illiquid bonds in

order to reduced those observed tracking errors.

3.2 ETF data

Specifically, we use an identification strategy with two main building blocks. First, we focus

on passively managed corporate bond ETFs newly launched during the 2007 and 2014 period.

Second, we require that the bonds included in the newly launched ETFs not be included in

an existing ETF. We do allow them to be held by existing bond mutual funds. Any empirical

results will be additional impact of ETF inclusion beyond that effected by the inclusion in

the existing bond mutual funds. However, if these bond mutual fund and ETFs are newly

launched at the same time, then we exclude the newly added bond from our treatment group

as both instruments improve the market accessibility and the interpretation of our empirical

results would be ambiguous.

To identify ETFs, we use the CRSP mutual fund database with the identifier, et flag.

To find corporate bond ETFs, we first limit the universe of ETF to the fixed income category,

that is, to those ETFs whose lipper asset code (lipper asset cd) is either TX or MB. We

exclude fixed income ETFs that concentrate on Treasury bonds or non-bond assets such as

commodities. From the remaining ETFs, we select those ETFs, in which corporate bonds

comprise at least 20% of the holdings. We identify the subset of the ETF constituent bonds

that are added to ETF portfolio for the first time within 6 months after the inception date

of ETFs.

A potential complication is that for some ETFs, their portfolio holdings were reported to

CRSP with a significant delay, as also noted in Dannhauser. This delayed reporting problem

is more prevalent with ETFs that were launched before 2007. To alleviate this concern, for

18

ETFs offered by BlackRock iShares, we hand collected the monthly portfolio holdings from

the funds’ websites. For other index ETFs with the reporting delay problem, we assume

that the identities of the portfolio holdings do not change significantly between the inception

date and the earliest report date, if made within a year. We take the holdings on the earliest

reported date to be the initial ETF portfolios.

Due to data availability on ETF holdings, our data starts from 2007, which is also the

year in which the number of fixed income ETFs and the aggregate assets under management

increased rapidly (Figure 4). Hence our data are sufficiently large to test the theoretical

predictions in Section 2. However, this will exclude the very first ETFs such as LQD (iBoxx

investment grade corporate bond ETF, launched in 2002) from our analysis.

Table 1 summarizes the 216 corporate bond ETF inceptions. These 216 new ETFs are

sponsored by 24 fund families in total, the largest eight of which are shown in Panel A of

Table 1. The very first corporate bond ETFs were offered by BlackRock in 2002. Corporate

bond ETF inceptions jumped in 2007 and have been spread out fairly evenly over time since

then. Restricting to funds in the CRSP portfolio holdings database reduces the number of

ETFs to 117 as in Panel B of Table 1. In particular, the holdings data for the early ETFs

from 2002 and 2003 are absent due to a long reporting delay. The inception of these 117

ETFs involves 106 unique fixed income securities per fund on average; the maximum is 1840.

On average, these inceptions include 90 securities that appear for the first time.

After the filtering of the bonds described in section 3.1, these 117 newly launched ETFs

are associated with 2,623 corporate bonds that are added to their portfolios for the first time

and whose trades are reported to TRACE (Figure 4). In our data sample, the corporate

bond transaction data starts from 2008 when FINRA started disseminating the buy and sell

indicators (Dick-Nielsen, 2009). Since 2010, ETF portfolios have become larger and a larger

number of bonds have been added.

Identification also requires that ETF introductions increase market accessibility for in-

vestors. To test whether they do, we examine changes in empirical proxies for market

19

accessibility such as fees, share classes, spreads around ETF inclusions. We also examine

characteristics of corporate bonds included in the newly launched ETF portfolios for the first

time. Throughout we use difference-in-difference tests to help remove secular trends that

affect similar bonds at the same time. The purpose of our diff-diff tests is to test whether

bonds that experience exogenous ETF inclusions undergo changes in liquidity that exceed

those expected in the absence of such inclusion.

To remove common influences, diff-diff tests compare the change in liquidity for treated

bonds to the contemporaneous change for a set of control bonds matched to have similar

characteristics but which are unaffected by the treatment. We match bonds on trading

volume before the inclusion in ETF, credit rating at the time of trading, time to maturity,

and issue size based on the prior literature. In addition, we match for mutual fund holdings

and dealers’ inventory positions. Specifically, we choose as controls for each included bond,

one bond based on the observed characteristics, subject to the condition that control bonds

did not experience a new mutual fund inclusion in the six months before or after.

3.3 Corporate bonds intraday trade data

In order to estimate the transaction costs of corporate bonds, we use the enhanced TRACE

intraday trading data for corporate bonds issued by U. S. firms as our primary database.

TRACE covers over 95% of U.S. corporate bonds traded since October 2004 (Rossi, 2014).11

The exact dollar transaction trading volumes are available. We use the masked dealer identity

information available in the enhanced TRACE and control for the dealer inventory changes

in relation to the bond liquidity as a consequence of ETF inclusion. Our dataset consists of

corporate bonds that were added to an ETF within 6 months of its inception date, but never

included in other ETF portfolios previously, and whose trading was reported to TRACE.

We exclude transactions associated with issuing, cancelled orders, sellers’ option settle-

ments, or reversals, following Dick-Nielsen (2009). Inter-dealer transactions may have been

11Corporate bonds traded in NYSE Automated Bond System (ABS) are not reported to TRACE.

20

reported twice by both of the dealers, in which case we keep only one trade. Large changes

of bond price during the day are likely to arise for reasons unrelated to liquidity pressure, for

instance, when an ETF manager possesses private information about the true value of the

bond or has skills to forecast future price changes of the bonds. We therefore remove bond

transactions whose prices at two consecutive time periods deviate more than 10%, following

Campbell and Taksler (2003).

We obtain historical bond credit ratings from DataStream and the issue amount, the issue

date and the issuer information from Bloomberg. For missing data, we use the Mergent Fixed

Income Securities Database (FISD) to complement the information. The three major credit

rating agencies – S&P, Moody’s, Fitch – have different rating categorizations. Following

the convention in the credit rating literature (Becker and Ivashina, 2015; Cornaggia and

Cornaggia, 2013; Bongaerts, Cremers, and Goetzmann, 2012), we assign a numeric score for

the credit rating scale from each rating agency (Becker and Ivashina, 2015) and merge the

scores for each bond as follows: At any date, if three rating scores are available, we take the

middle score. If two rating scores are available, we take the lower score. As the credit rating

upgrades or downgrades for each bond, this combined rating score is adjusted.

We also use a separate database for the 144A bond transactions in TRACE (trace btds144a)

and investigate the liquidity spillover between ETFs and the underlying bond market in sec-

tion 4.3.

3.4 Public information variables

Two of the public information variables – on-the-run Treasury security returns and the yield

spread between BAA bonds and the risk free rate – are obtained from the Federal Reserve

Statistical Release. The third, the percentage returns on the issuing firms’ common stock,

is obtained from the CRSP daily stock returns database.

21

3.5 Corporate Bond Liquidity Measure

Many of the studies on intraday trade execution costs in equity markets rely on pre-trade

bid and ask quotes and construct measures such as quoted spreads or effective spreads.

Alternatively, daily trading cost of securities can also be estimated from daily returns. For

instance, Lesmond, Ogden, and Trzcinka (1999) developed the LOT measure, which has

been applied to the global market by Fong, Holden, and Trzcinka (2017). It assumes that

informed investors trade only when their expected profits outweigh the transaction costs of

the securities. The daily returns themselves implicitly reflect trading costs, which can be

recovered computationally. Chen, Lesmond, and Wei (2007) applied the LOT measure to

corporate bonds and showed that it is correlated to their available quotes.

In contrast to equity market, pre-trade quotes for the majority of corporate bonds are

not available and hence effective or quoted spreads are not viable measures.12 To overcome

this limitation, two alternative measures have been used in the literature. The first method

is estimating transaction costs of corporate bonds by investigating a subset of the bonds

that have both buy and sell orders on each day. The difference between the average buy

and sell prices is used as a measure of the daily transaction costs of the bond. Since it is

common for corporate bonds to have only buy trades or only sell trades on a given day, this

method significantly reduces the number of usable data. Inferring the transaction costs for

other non-traded bonds from it could potentially be limited (Schultz, 2001).

A second method is the indicator-based regression model, which was originally suggested

for equity markets by Huang and Stoll (1997), and then extended to corporate bonds by

Bessembinder et al. (2006). This method decomposes the spreads into adverse selection

costs, inventory costs, and order processing costs. This model does not require that bond

trades are present both on buy and sell sides and is applicable as long as some trades occur.

12Pre-trade quotes are available for corporate bonds traded in NYSE Automated Bond Systems, but arenot reported to TRACE. Reuters Fixed Income Database provides the daily quotes from major bond dealers.Recently, Schestag, Schuster, and Uhrig-Homburg (2016) use the daily bond prices reported in Datastreamand measure the liquidity of bonds.

22

However, very few trades in a short time period or very sparsely traded bonds will make the

cost estimates noisy.

In this paper, we measure corporate bond liquidity by adopting the indicator-based re-

gression model, which consists of two steps. In the first stage, we estimate the serial cor-

relation among the order flows Qt using ordinary linear regression (OLS), Qt = ρQt−1 + εt.

The estimated unexpected order flow is ε̂t = Qt − ρ̂Qt−1, which will be used in the sec-

ond stage. We let Q∗t = ε̂t. In the second stage, we investigate the changes of transaction

costs of bonds with their inclusion in an ETF. To account for a potential parallel trend

present in the bond market, we extend the main model in Bessembinder et al. (2006) to the

difference-in-difference framework,

∆Pt = a+ α1S∆Qt × ETF × Post

+ α2S∆Qt × ETF + α3S∆Qt × Post+ α4S∆Qt + γSQ∗t + wXt + δt. (4)

ETF is a binary variable that takes 1 if the bond is included in ETF and 0 otherwise. Post

is a binary variable that takes 1 if the bond is traded after the inception date of ETF and

0 if the trade occurs prior to the first offer date of the ETF. The coefficient α1S is our

main interest and measures changes of transaction costs of ETF bonds relative to those for

non-ETF bonds, due to the inclusion in an ETF.

Following Bessembinder et al. (2006), Xt includes three components representing public

information available at the time of trading. The first is the yield, on the bond transaction

day, of the on-the-run Treasury security matched for time to maturity with the corporate

bond. Since the maturities of corporate bonds vary, this Treasury yield variable is indica-

tive of changes in the interest rate term structures. The second is the daily default spread

as the difference between the yield rate of aggregate BAA bonds and the risk free rate.

As market sentiment leans negatively (or positively), the risk premium for aggregate BAA

bonds increases (decreases). Hence, this default spread allows us to control for market wide

23

perceptions on the overall economy. The third is the return of the common stock for the

issuing firm since the last transaction date of the bond. All bonds issued by subsidiaries

are considered to have the same parent company. The return on the common stock reflects

firm-specific news. For instance, a positive earnings announcement increases investors’ ex-

pectation of their future earnings, resulting in increases both in the firm’s stock price and

return.

We estimate the second stage regression with the weighted least squared regression (WLS).

Since recent trades are likely to be more serially correlated with today’s trades than are dis-

tant trades, we use as weights the inverse of the number of days elapsed since the last

transaction day. For instance, if the last trade was made two days ago, then all the trades

for today are given the equal weight of 1/2.

To select the non-ETF bonds in the control group, we match on the following charac-

teristics of the bonds, using the propensity score matching algorithm (Wooldridge, 2010):

trading volume before the inclusion in ETF, credit rating at the time of trading, issue size,

age, time to maturity of the bond at the time of trading, ownership by bond mutual funds,

and dealers’ inventory capital. Prior studies (Schultz, 2001; Edwards et al., 2007; Goldstein,

Hotchkiss, and Sirri, 2007; Hendershott and Madhavan, 2015) showed that credit rating and

trading volume are robust explanatory variables for the liquidity of the bonds: bonds with

higher credit rating or with large trading volume are less costly to trade. We include is-

sue size because large issues are more frequently traded due to their availability. Trading

volume also influences the spread because it allows bond dealers to manage their inventory

better (Grossman and Miller, 1987). The time to maturity is relevant to the bond liquidity

because long term bonds are harder to trade in general. The age of the bond matters because

the majority of corporate bonds are known to end up being owned by institutional investors,

who infrequently trade, within two years of issuance. Furthermore, bond ownership by

buyside institutions are important as recent evidences such as in Anand, Jotikasthira, and

Venkataraman (2018) show that they also provide liquidity in bond markets. Therefore, in

24

the pre-ETF period, the two groups of bonds are similar in their liquidity. We are interested

in whether their transaction costs of the bonds diverge in the post-ETF period.

We note that the standard statistical inference with the t-statistic is not valid for our

analysis for the following reasons. The second stage regression has an explanatory variable

Q∗t = ε̂t, which was estimated from the first stage. The correct inference for α1S must include

the adjustment taking into account possible interactions of the noise terms between the first

and second stage regressions (Wooldridge, 2010). Furthermore, some bonds are traded more

than once within a day. This implies that we have multiple observations for the realized

∆Pt for the time t. The number of observations per day is random. Also, if some bonds are

issued by the same parent firm, or their times to maturities are similar, then changes of bond

prices are likely to be correlated. Therefore, the noise term in the second stage regression is

less likely to form the standard t-distribution.

In order to estimate the true distribution of the noise term, δt, we use the block boot-

strapping method. For each bond, since the daily trades are more likely to be correlated

than inter-day trades, we sample with replacement trades from daily trades 100 times. The

correlation structure between the days remains the same. But at each time, the number of

trades occurring is randomly varying. The empirical distribution for the bond’s transaction

cost α̂1S approximates the true population distribution (Wooldridge, 2010). We use this

empirical distribution and test against the null hypothesis that the true transaction cost is

zero, i. e. H0 : α1S = 0. We compute the p-value as the likelihood that the estimate α̂1S

has the sign opposite of the fitted value, given the null hypothesis.

4 The Effect of ETF Inclusion on Bond Liquidity

4.1 Does ETF inclusion increase trading volume?

The theoretical predictions in section 2 imply that inclusion in an ETF portfolio increases

trading volume of the underlying bonds. To examine whether it does, we investigate changes

25

of trading volume around ETF inclusions. In addition, we use difference-in-difference (diff-in-

diff) tests to account for secular trends that affect similar bonds at the same time. The goal

of diff-in-diff analysis is to test whether bonds that are included in ETF portfolios experience

changes of trading volume that exceed those expected in the absence of such inclusion.

To account for contemporaneous trends, our diff-in-diff tests contrast the change in trad-

ing volume of treated bonds to that of control bonds matched to have similar characteristics.

We choose a set of control bonds that were not included in any ETF portfolios during our

sample period, but whose trades were reported to TRACE. We match bonds on credit rat-

ing, issue size, age and time to maturity on the inception date of the ETF, and cumulative

trading dollar volume six months prior to inclusion in the ETF, using the propensity score

algorithm. More specifically, for each treated bond, we choose as a control one bond that

is the closest to the treated bond in terms of the relevant liquidity measure – credit rating,

issue size, age, time to maturity, and cumulative trading volume in the preceding six months

of the ETF inception date.13 These control bonds were not included in any ETF portfolio

in our sample period before or after.

Table 2 describes static characteristics of both ETF and non-ETF bonds in the matched

sample. About 55% of the bonds have the dollar issue size larger than $500M. Bonds with

superior credit quality (AA− and up) compose 12% of the sample. About 63% of the bonds

have investment credit rating of BBB− through A+. Non-investment grade bonds comprise

26% of the sample in both of the groups.

Trading characteristics of the bonds are shown in Table 3. The bonds are traded close

to the par values. The aggregate dollar trading volume decreases slightly for both treatment

and control groups, which is consistent with the findings of Randall (2015) that the overall

trade size of corporate bonds has decreased since 2007. While the total dollar trading volume

of ETF bonds decreases, it increases for low volume, small trade size ETF bonds. The total

number of trades increases for ETF bonds, while it decreases for non-ETF bonds. These

13We also conducted the analysis by selecting multiple bonds with similar characteristics as controls. Theresults remain the same. We report only the results where We choose one matching bond.

26

divergent changes are most prominent for small trade sizes less than $1M.

As a preliminary analysis, we also compare changes in transaction costs of the bonds using

the two methods discussed in section 3.5. Comparison of same-day buy and sell transactions

does not show a significant change in transaction costs, both for ETF bonds and non-ETF

bonds. However this measure is noisy due to the limited sample size. The effective half-

spread using the indicator-based regression method without any control variables does show

a significant decrease in transaction costs. For ETF bonds, the smallest quintile based on

the cumulative trading volume experiences the greatest decrease – about 30% in transaction

cost or 15 basis points. Non-ETF bonds show an increase in transaction costs.

This evidence is consistent with the prediction of the model that inclusion in an ETF

portfolio decreases transaction costs. The benefits of inclusion accrue most to the underlying

bonds with smallest trading volume. Furthermore, Figure 5 shows the parallel trends of the

bonds both for the treatment and control groups. For high yield bonds, the parallel trends

diverge after the inclusion in ETFs. The liquidity of high yield bonds improves significantly,

while that of investment grade bonds remain close to each other.

4.2 Bond liquidity improvement following ETF inclusion

4.2.1 Full sample analysis. To test Proposition 1 more formally, we estimate changes

in transaction costs for bonds included in an ETF for the first time using the indicator-based

regression model in equation (4) which includes all control variables. The coefficients are in

percentages. Table 4 reports the results. The most interesting estimated coefficient is the

one on ∆Q× Post× ETF.

We estimate that the average effective half spread for trading individual corporate bonds

is 42 basis points during our sample period. This is consistent with the findings of Bessem-

binder et al. (2017) that, amid the concern on the overall bond market liquidity after the

financial crisis, the bond liquidity gradually recovered back to its previous level from 63 bps

for the 2009–2010 period to 47 bps for the 2010–2014 period. In addition, this confirms

27

that our bond data samples are mostly traded in the dealers’ market, as Hendershott and

Madhavan (2015) estimated that cost of trading odd lots ($100K∼1M) for corporate bonds

in 2011 was 48 basis points for the dealers’ market and 14 basis points for the electronic

trading platform. Furthermore, in accordance with Proposition 1, the liquidity of corporate

bonds improves significantly following ETF inclusion by 3.5% or 1.3 bps on average, after

taking into account changes in the control bonds.

To test the estimated liquidity improvement of the treated bonds for statistical signifi-

cance, we use the block-boostrapping method to address potential bias due to correlations.

There are two potential souces of bias. When an ETF is newly launched, multiple corpo-

rate bonds are added to the ETF portfolio, which could imply that these bonds are cross-

sectionally correlated in their liquidity. Furthermore, bond price changes are referenced

relative to the last prior-day transaction price. If multiple bond transactions are found in

one day, then they are likely to be correlated in time. Following the standard in the lit-

erature (Wooldridge, 2010), we adopt the block-bootstrapping method while keeping the

correlation matrix the same. Our results are highly statistically significantly different from

zero.

In addition, we find that the coefficient of Qt−1 is positive and statistically significantly

different from zero. This implies that bond orders are positively serially correlated. The

coefficient of the unexpected order flow Q∗t is also positive significant. Since this coefficient

indicates the informational content observed in the order flows of the bonds, this result

implies that some bonds are traded on private information.

The estimated coefficients of the public information variables are significant. Stock re-

turns of the issuing firm have a positive impact on the bond transaction costs. The impact

is larger for non-investment grade bonds than for investment grade bonds. This is consistent

with the finding of Hotchkiss and Ronen (2002) that both stock and bond returns respond to

new information about the value of the issuing firm’s underlying assets. The new information

is more valuable for non-investment grade bonds, because the investors and the firms are

28

more likely to be asymmetrically informed. The coefficient estimates for the default spread

is negative and significant for investment grade bonds, and positive and significant for the

non-investment grade bonds.

4.2.2 Alternative identification strategy: Quasi-Natural Experiment for a high

yield bond ETF. To provide additional support for the causality of our results, we explore

a quasi-natural experiment, where the reason for the ETF portfolio inclusion is manifest.

The same experiment has previously been used by Dannhauser (2017) to show the valuation

effect of corporate bond ETFs. The iShares iBoxx $ High Yield Corporate Bond ETF (HYG)

started trading on April 4, 2007 and is one of the largest high yield bond ETFs by assets

under management – its assets had already reached $3.3 billion as of June 30, 2009. The

HYG portfolio holds about 94% of its assets in non-investment grade corporate bonds and

the remainder in cash. The expense ratio is 0.5 basis points and the portfolio turnover ratio

is about 30%.

HYG is passively managed and follows the aggregate bond index managed by Markit for

their benchmark index. On June 22, 2009, Markit changed the index composition rule by

imposing a 3% cap on the constituent bonds and requiring the index to be computed as a

value-weighted average, effective June 30, 2009. As a consequence, during the second half

of 2009, about 248 bonds were newly added to HYG. While the governing board of HYG

may have known of the change in advance, and may have initiated trading of the relevant

bonds early, these types of trading activities are likely to have been limited due to concerns

about tracking errors. As indirect evidence, the number of newly added bonds increased

dramatically after June 2009: Only 16 were added from January to May, while 31 were

added in June, and 248 were added from July to December. Hence, it seems reasonable that

these bonds were added to HYG due to the rule change rather than to anticipated reduction

in transaction costs. We investigate changes of the liquidity of these bonds using a diff-in-diff

framework and consistently find that ETF inclusion led to a reduction of transaction costs

29

of the bonds by 3.5%. (Table 10)

The fact that we find bond liquidity improves following inclusion in ETF portfolios us-

ing two independent reasons for inclusion – 117 newly launched ETFs and HYG whose

benchmark index underwent a change in the composition rule – supports our predictions

and identification strategies. While we cannot completely rule out a possibility that ETF

managers might have timed the timing of the funds’ inception date in a way that would

spuriously lead to our results, it seems apparent that the bond inclusion in HYG during the

later 6 months of 2009 were involuntary.

4.2.3 ETF arbitrage as the underlying mechanism. As evidence for the underlying

economic mechanism of our results, we estimate ETF arbitrage and investigate whether the

liquidity spillovers are in fact facilitated by arbitrageurs. ETF shares can be created or

redeemed by authorized participants (AP) by exchanging a basket of the securities for ETF

shares. When the net asset value of the ETF portfolio diverges from ETF market prices,

the APs wish to take advantage of these arbitrage opportunities through the creation and

redemption process. Therefore, we measure as a proxy for ETF arbitrage the cumulative

changes in the number of ETF shares outstanding relative to the average ETF shares out-

standing during the preceding 6 months.14 If this ETF arbitrage proxy exceeds 1.25, we

consider the associated ETFs to have high ETF arbitrage. In agreement with our predic-

tions in Proposition 1, we find that the bond liquidity improves only when ETF arbitrage

are high (Table 5). Conversely, when ETF arbitrage is low, the bond liquidity deteriorates.

Furthermore, a potential concern is that ETF arbitrage may become impaired when

the required conditions for ETF arbitrage are not present simultaneously. For instance,

14This ETF arbitrage proxy is likely to be a lower bound on true arbitrage for the following reasons. First,ETF arbitrage can occur entirely in the secondary market without having to create or redeem ETF shares.Second, as Evans, Moussawi, Pagano, and Sedunov (2017) show, not all of ETF creations and redemptionsare completed. Often ETF shares are sold short first, but the underlying securities fail to be delivered withinthe T + 3, up to T + 6, settlement days (“operational-shorting”). Since only the completed creation andredemptions change ETF shares outstanding, our ETF arbitrage proxy is likely to underestimate the truearbitrage. In spite of our conservative estimate of ETF arbitrage, we find that the liquidity spillover by ETFarbitrage are still present.

30

while differences between ETF market prices and the net asset values (NAV) of the ETF

portfolio would signal arbitrage opportunities, a dealers’ inventory imbalance might limit

their arbitrage activities (Pan and Zeng (2017)). To investigate the impact of these two

requirements on ETF arbitrage, we measure as ETF arbitrage opportunities daily differences

of ETF market price and its NAV. To measure the dealers’ inventory imbalance, we calculate

differences of the buy and sell dollar volumes of all the constituent bonds of the ETF portfolio

on each day. We then partition the data into four groups of trading days depending on

arbitrage opportunities and the dealer inventory imbalance (Table 6). We find that the

liquidity spillover from ETFs to the underlying bonds prevails in all four groups. However,

since we require ETFs to have the daily order imbalance and each group relies on a subset

of the total sample days, the sample size is smaller than that for the full analysis.

4.2.4 Interpretation. Our results do not exclude the possibility that the liquidity spillover

mechanism through ETF arbitrage could potentially be fragile at times, especially in the pri-

mary market. The intuition is that ETF authorized participants play a dual role as ETF

arbitrageurs and also as bond dealers. Under the conditions when their inventory is con-

strained, there could be a temporary limit to arbitrage between the ETF and the underlying

corporate bonds, which might hurt the liquidity of the underlying bonds (Pan and Zeng,

2017). The interpretation of our results is that the overall effect of ETF trading in the

long run benefits the liquidity of the underlying bonds due to the effect of the secondary

market. A more recent, but preliminary, analysis suggests that whenever ETF arbitrage

opportunities are high, the trading volume of the underlying bonds increases.

4.2.5 Cross sectional variation of bond liquidity improvement. The effect of ETF

portfolio inclusion on liquidity improvement is likely to vary among bonds that have the

same minimum trading size, but whose market accessibility varies along other dimensions

such as information asymmetry and unfamiliarity. As Corollary 1 predicts, the liquidity

improves more as the trading barrier prior to the introduction of ETFs was higher. Prior

31

evidence suggests that bond market segments with small trading volume, lower credit rating,

or long-term maturity are harder to access for uninformed investors.

Table 7 reports that the underlying bond trading cost prior to the initial trading of

ETF is linearly related to the cumulative trading volume, which is consistent with earlier

findings (Edwards et al. (2007), for instance). The bonds with the smallest dollar trading

volume (VOL 1) have highest trading cost at 92 basis points. As the trading volume increases,

the transaction costs decrease. The most liquid bonds with large dollar trading volume have

lowest trading cost at 41 basis points.

The initiation of ETF trading results in differential changes of trading costs of the under-

lying bonds. The changes are linearly related to the cumulative trading volume. The illiquid

bonds (VOL 1) become less costly to trade and their transaction costs decrease by 9 basis

points, i. e. about a 10% reduction. On the other hand, liquid bonds with median trading

volume become more costly, increasing the trading costs by 2 basis points. The most liquid

bonds (VOL 5) show little impact on the liquidity due to ETF trading.

Table 8 reports transaction cost estimates arranged by the bond credit rating. Before

the introduction of ETFs, non-investment grade bonds are more costly to trade at an esti-

mated 57 basis points, while investment grade bonds cost 37 basis points. This difference

of trading cost can be explained by large information asymmetry between the issuing firm

of non-investment grade bonds and the investors (Holden, Mao, and Nam, 2016). With the

introduction of ETF, the ETF trading from the uninformed investors has an indirect impact

in the underlying junk bond market. The effective spreads of the junk bonds held by ETFs

decrease by 10%, i. e. 5.4 basis point.15 This is consistent with our theoretical predictions

in Proposition 1.16

15As discussed earlier in the introduction, Dannhauser (2017) finds that the relation between ETF tradingactivities and the underlying bond liquidity is not robust and that ETF ownership is correlated with liquiditydeterioration only for investment grade bonds. This is not in direct conflict with our results due to differencesin methodology. We focus on event study with six month window before and after ETF inclusion, whileDannhauser looks at a long panel study over four years with ETF ownership.

16Strategic inventory management by the bond dealers might provide an alternative explanation of ourcross-sectional results. Recent literature (Bessembinder et al. (2017); Choi and Huh (2017); Goldstein andHotchkiss (2017), for instance) has shown that corporate bond dealers manage their inventories strategically

32

Unlike bonds, fixed income ETFs have no maturity. As some incumbent bonds near

maturity, ETF managers may wish to sell off the maturing bonds in order to reduce the

tax burden from realizing the capital gains. This specific trading demand for these bonds

may offset the positive impact of ETFs on improving the liquidity of the underlying bonds.

In Table 9, we show that with the introduction of ETFs, the liquidity improves for the

bonds whose time to maturity is more than two years, while the liquidity deteriorates for

the maturing bonds. ? finds similar results by extrapolating intraday bond quotes from the

Bloomberg quotes available at the end of previous day as an alternative measure of bond

liquidity, which indicates that our results are robust with different definitions of liquidity.

4.2.6 Multivariate Analysis. While all our results so far are univariate, potential cor-

relations among the explanatory variables may weaken our primary results. To alleviate

such concern, in Table 12, we run a multivariate analysis by combining all the explanatory

variables. Our univariate analysis results continue to hold. In addition, as in column (1),

bond trading volume is the dominant factor in determining the transaction cost of the bonds.

The low volume bonds benefit most from the inception of ETF trading. The bonds whose

remaining life is more than 2 years had lower transaction costs by 31 basis points than those

with little time to maturity. With the inception of ETF trading, these bonds near maturity

exhibit an even bigger transaction cost by 2.6 basis points. In columns (2) and (3), we con-

ducted the multivariate analysis based on the bond credit rating. The univariate analysis

from Table 8 remain unchanged.

4.3 The impact of ETF trading on 144A bonds

Liquidity spillover is likely to have the largest effect on 144A bonds, which were offered only

to qualified institutional investors (QIB) until ETFs began to increase their holdings of 144A

by closing out their positions for high yield bonds sooner, with a concern for larger adverse selection andhigher risk capital requirement. However, to the extent that this strategic market making is invariantaround the event of ETF inclusion, this result - the liquidity improvement of the underlying securities - isstill interpreted as an impact of the inclusion in ETF portfolios.

33

corporate bonds in 2010 (Figure 4(B)). One reason is that, since non-institutional investors

are not legally allowed to trade 144A bonds, liquidity reduction due to discretionary investors

leaving the underlying market is limited. More importantly, after the inception of ETFs,

discretionary investors can indirectly access the 144A bond market. As trading volume of

such ETFs increases, the liquidity of these restricted securities is likely to improve. We

empirically test this prediction in this section.

While 144A bonds are restricted securities, they have undergone a series of regulatory

changes that have allowed ETFs to trade them. Unlike public debt, which requires regis-

tration and SEC oversight, firms can privately issue debt if they place it with accredited

investors or a limited number of individual investors who intend to hold it for investment.

This privately placed debt can be resold if the issue is registered with SEC after issuance or

if investors have held it for, typically, at least two years. In 1990, Rule 144A was adopted

to allow investors to immediately resell 144A securities without SEC registration or the two

year holding period, as long as they are sold to qualified institutional buyers (QIB). On

August 15, 2007, NASDAQ offered an electronic trading platform – PORTAL – for these

privately placed securities, the access to which was available only to QIBs and the data of

which were not made available to the public.

Furthermore, in accordance with the Jumpstart Our Business Startups Act (JOBS Act)

section 201(a)(1), SEC finalized a rule in 2012 August that permits the sale of 144A bonds to

persons, other than QIBs, who QIBs believe are QIBs.17 This implied that non-QIBs such as

ETFs could easily trade 144A securities for the first time. As a consequence, the size of 144A

bond market has become large. During the year 2012, firms raised overall capital of $1.2

trillion through regular debt and equity offerings, while $0.636 trillion worth of capital was

raised through 144A offerings.18 To facilitate trading, FINRA started disseminating trans-

action data for 144A bonds as well as certain other private placements.19 As a consequence,

17See https://www.sec.gov/rules/final/2013/33-9415.pdf18See https://www.sec.gov/rules/final/2013/33-9415.pdf19See SEC Release, No. 70345, 16 September 2013. Also, https://www.finra.org/newsroom/2014/finra-

brings-144a-corporate-debt-transactions-light

34

during the third quarter of 2014, transactions of 144A securities exploded, comprising nearly

9% of the average daily volume in investment-grade corporate debt, and nearly 28% of the

average daily volume in high-yield corporate debt (Table 13).

Although 144A securities are liable to be illiquid due to the required holding periods

before resale, they have been popular among debt issuers and institutional investors. U.S.

and foreign issuers take advantage of the speedy issuance in the 144A market.20 As of 2013,

these private placements comprise about 15% of high yield bond issuance (Table 13).21 In

addition, these 144A securities have lower transaction costs than those that are otherwise

equivalent, because trades occur between institutional investors and hence adverse selection

cost is low (Edwards et al., 2007).

Therefore, the increased holdings of the 144A bonds by ETFs likely provide further liq-

uidity in the underlying bond market with limited access. We exploit the TRACE 144A

bond database that has recently become publicly available. Consistent with our theoretical

predictions made earlier in Section 2, we find that with the introduction of ETFs, the trans-

action cost of the 144A bonds owned by ETFs decreases significantly by 5 basis point (Table

14). The reduction of transaction costs is larger for large volume bonds, for high-yield bonds,

and for long-term bonds (not shown).

4.4 Robustness

Since we match the bonds in the treatment group with those in the control group by using

the bond characteristics, one potential concern would be that our results may be driven by

unobservable characteristics of the bonds that are related to their issuing firms. In order

to address this concern, we added firm fixed effects in our analysis. Our results remain the

20Fenn (2000) showed that about 97% of the 144A high-yield securities are subsequently registered withthe SEC. Among high yield 144A debt, 58% was issued by first-time bond issuers, while 29% was issued byprivately owned firms. The credit risk premium for 144A high yield securities is higher than for public debt,but it gradually disappears after issuance. (Chaplisky and Ramchand, 2004)

21In an unreported note, we find that the composition of these privately placed debt is even higher amongthe most risky bonds whose credit rating is below B. As of 2013, they comprise about 50% of the issuanceof the most risky bonds.

35

same (Table 15).

4.5 Discussion

A tenable reading of our results is that retail investors value better diversified portfolios

and react to the newly available exposure to the corporate bond market provided by ETFs.

When individual corporate bonds are included in ETF portfolios, their liquidity improves

as a result. In principle, the exclusion of corporate bonds from any ETF portfolio should

have the opposite effect as exclusion removes the benefit of the liquidity spillover from ETFs.

This prediction has been confirmed by the recent study by Ye (2019).

5 Conclusion

Multi-market trading models often assume that investor market participation is homoge-

neous, but our empirical observation is that some markets are easier to access than others.

How important heterogeneous market accessibility ex ante is in determining response to in-

novation in the financial market is not well understood. This paper provides theoretical and

empirical evidence that market participation plays an important role in how liquidity of the

underlying securities changes when basket securities are introduced. To do so, we exploit

corporate bond ETF inceptions as the events that lower trading barriers.

We first extend the Kyle model to multi-markets where the degree of accessibility varies.

The discretionary investors are uninformed and concerned about having a tradable diversified

portfolio. We show that they prefer to invest in basket securities because of (1) a low expected

loss of trading and (2) ease of accessibility (i.e. lower minimum investment requirement). In

particular, the ease of accessibility in the basket market is critical in order to improve the

liquidity of the underlying securities.

As empirical evidence, first, we exploit TRACE bond transaction data for OTC corporate

bonds and show that in contrast to stocks, trading bond ETFs reduces the transaction cost of

36

the underlying bonds. The reduction of transaction costs is larger for low volume, high-yield,

and long-term bonds, to which the access was limited to retail investors. On the other hand,

transaction costs for high volume, investment grade, or short-term bonds increase. Second,

144A bonds, which were previously inaccessible to non-institutional investors, experience

improved liquidity as ETFs increase their holdings of 144A bonds. The aggregate savings in

trading an ETF bond is as big as −$18.5 million during the six months after inception of a

fixed income ETF containing that bond.22

Our results imply that bond investors benefit from the inclusion of bonds in fixed income

ETFs. This benefit accrues more to small retail bond investors and to junk bond investors,

to which the accessibility is limited previously. We believe the mechanism we proposed in

this paper is general and could potentially be extended in the future to hard-to-trade assets

such as mortgages or hard-to-access markets such as frontier markets.

22=10.259× (−0.0881)+34.897× (−0.0355)+78.447×0.0185+155.002× (−0.0015)+844.282× (−0.0011)=−1.85. This is equivalent to 18.5 million dollars.

37

Appendix A

A.1 Unconstrained solution

In the pre-ETF period, we solve the portfolio trading allocation problem using a constrained

optimization method. The Lagrangian of the objective function is set up with multipliers,

η1 and η2 as

L(α1, α2) = E(W )− γ

2Var(W )− η1 [E(|α1lqP1|)−B1]− η2 [E(|α2lqP2|)−B2] ,

subject to the following conditions

η1 [E(|α1lqP1|)−B1] = 0, η2 [E(|α2lqP2|)−B2] = 0

E(|α1lqP1|) ≥ B1, and E(|α2lqP2|) ≥ B2,

η1 ≥ 0, η2 ≥ 0.

We first look for unconstrained solutions by assuming the minimum trading requirement

Bi is sufficiently small. The constraints E(|αilqPi|) ≥ Bi (i = 1, 2) hold for all q, which

implies that the Lagrange multipliers ηi is zero. Solving the first order condition (FOC) of

the Lagrangian L(α1, α2) above, we obtain

α?i =1

γ

(W0 −

σ2l

Ai+ γ

σ4l

Ai

)[W 2

0 σ2i + µ2

iσ2l (σ

2i + 1)− σ2

l µ2iσ

2i

(k2i + ki + 1

k2i + 2ki + 1

)]−1

,

where Ai =[

ki(ki+1)2

µ2i σ2i

σ2u+σ

2l

]−1/2

.

In the post-ETF period, the discretionary uninformed investors have a expanded oppor-

tunity set for trading. The Lagrangian of the objective function is found as follows.

L(α1, α2, αe) = E(W )− γ

2Var(W )− η1 [E(|α1lqP1|)−B1]− η2 [E(|α2lqP2|)−B2] ,

38

where W =(α1W0

µ1+ α1lq

)v1 +

(α2W0

µ2+ α2lq

)v2 +

(αeW0

µe+ αelq

)ve − (P1 − v1)α1lq − (P2 −

v2)α2lq − (Pe − ve)αelq, depending also on the value ve and price Pe of the ETF. We assume

the ETF market is highly accessible by assuming the minimum trading requirement of ETF

is zero. Hence, the auxiliary conditions remain the same as above.

We let αi = α?i − αewi for i = 1, 2. With these αi, E(W ) and Var(W ) will be the same

as those in the pre-ETF period. To further increase the investor’s utility, we minimize the

loss due to trading by adjusting αe. We let the optimal trading demand of the ETF be

α̃?e. This α̃?e is strictly positive. The optimal trading demand for each security is therefore

α̃?i = α?i − α̃?ewi and is strictly smaller than α?i .

A.2 Constrained solution

Without loss of generality, we focus on the first asset market. In the pre-ETF period, the

optimal trading solution α1 binds with the constraint, E(|α1lqP1|) = B1. Solving for α1,

B1 = E|α1lqP1| = E|α1lq(µ1 + λ1(α1lq + u1 + x1))|

= E|α1||µ1||lq|+ |λ1|α21E(l2q) + E|α1lq||x1 + u1|.

Note that the third term is zero, because lq and x1 + u1 are uncorrelated random variables

with Normal distributions. In addition, E(x1 +u1) = 0. The two variables lq and x1 +u1 are

independent and so are |lq| and |u1+x1|. We also note that |α1λ1| = α1λ1 =√A1/

√σ2u + σ2

l ,

where A1 = σ21k1/(k1 + 1)2. Therefore, we find that

|α1| =B1

|µ1|[σl

√2π

exp(− µ2l

2σ2l

)+ µl

(1− 2Φ

(−µlσl

))]+√

A1

σ2u+σ

2l(σ2

u + σ2l ),

where Φ(·) is the normal cumulative distribution function. For some combinations of µl, σl, σu,

this α1 does not satisfy the FOC. If these infeasible trading solutions happen for all the dis-

cretionary uninformed investors, we call this market inaccessible. We assume these investors

39

trade a risk free security at that time.

On the other hand, in the post-ETF period, as the ETF market is highly accessible, these

uninformed investors reconsider their trading decisions. Their optimal trading demand αe

is strictly positive. The influx of these uninformed investors to the ETF market increases

ETF liquidity, which spills over the underlying market through arbitrage.

40

References

Abudy, M., and A. Wohl. 2017. Corporate Bond Trading on a Limit Order Book Exchange.

Review of Finance 22:1413–1440.

Anand, A., C. Jotikasthira, and K. Venkataraman. 2018. Do Buyside Institutions Supply

Liquidity in Bond Markets? Evidence from Mutual Funds. Southern Methodist University

Syracuse University Working Paper .

Becker, B., and V. Ivashina. 2015. Reaching for Yield in the Bond Market. Journal of

Finance 70:1863–1902.

Ben-David, I., F. Franzoni, and R. Moussawi. 2018. Do ETFs Increase Volatility? Journal

of Finance 73:2471–2535.

Bessembinder, H., S. Jacobsen, W. Maxwell, and K. Venkataraman. 2017. Capital Commit-

ment and Illiquidity in Corporate Bonds. Forthcoming in Journal of Finance .

Bessembinder, H., W. Maxwell, and K. Venkataraman. 2006. Market transparency, liquidity

externalities, and institutional trading costs in corporate bonds. Journal of Financial

Economics 82:251–288.

Bhattacharya, A., and M. O’Hara. 2016. Can ETFs Increase Market Fragility? Effect of

Information Linkages in ETF Markets. Cornell University Working Paper .

Biais, B., and R. Green. 2019. The Microstructure of the Bond Market in the 20th Century.

Forthcoming in the Review of Economic Dynamics .

Bongaerts, D., M. Cremers, and W. N. Goetzmann. 2012. Tiebreaker: Certification and

Multiple Credit Ratings. Journal of Finance 67:113–152.

Campbell, J., and G. Taksler. 2003. Equity Volatility and Corporate Bond Yields. Journal

of Finance 58:2321–2349.

41

Carey, M., S. Prowse, J. Rea, and G. Udell. 1993. The Economics of the Private Placement

Market. The Staff Report by the Board of Governors pp. 1–111.

Chang, C., and X. Yu. 2010. Informational Efficiency and Liquidity Premium as the Determi-

nants of Capital Structure. Journal of Financial and Quantitative Analaysis 45:401–440.

Chaplisky, S., and L. Ramchand. 2004. The Borrowing Costs of International Issuers: SEC

Rule 144A. Journal of Business 77:1073–1097.

Chen, L., D. A. Lesmond, and J. Wei. 2007. Corporate Yield Spreads and Bond Liquidity.

Journal of Finance 62:119–149.

Chen, Y., W. Ferson, and H. Peters. 2010. Measuring the timing ability and performance of

bond mutual funds. Journal of Financial Economics 98:72–89.

Choi, J., and Y. Huh. 2017. Customer Liquidity Provision: Implications for Corporate Bond

Transaction Costs. University of Illinois at Urbana-Champaign Working Paper .

Cici, G., and S. Gibson. 2012. The Performance of Corporate Bond Mutual Funds: Evidence

Based on Security-Level Holdings. The Journal of Financial and Quantitative Analysis

47:159–178.

Cornaggia, J., and K. J. Cornaggia. 2013. Estimating the Costs of Issuer-paid Credit Ratings.

Review of Financial Studies 26:2229–2269.

Da, Z., and S. Shive. 2018. Exchange traded funds and asset return correlations. European

Financial Management 24:136–168.

Dannhauser, C. D. 2017. The Impact of Innovation: Evidence from Corporate Bond ETFs.

Journal of Financial Economics 125:417–648.

Dick-Nielsen, J. 2009. Liquidity Bias in TRACE. Journal of Fixed Income 19:43–55.

42

Edwards, A. K., L. E. Harris, and M. S. Piwowar. 2007. Corporate Bond Market Transaction

Costs and Transparency. Journal of Finance 62:1421–1451.

Ellul, A., P. Jotikasthira, and C. T. Lundblad. 2011. Regulatory pressure and fire sales in

the corporate bond market. Journal of Financial Economics 101:596–620.

Evans, R. B., R. Moussawi, M. S. Pagano, and J. Sedunov. 2017. ETF Short Interest and

Failures-to-Deliver: Nakes Short-selling or Operational Shorting? University of Virginia

and Villanova University Working Paper .

Fenn, G. W. 2000. Speed of issuance and the adequacy of disclosure in the 144A high-yield

debt market. Journal of Financial Economics 56:383–405.

Fong, K. Y. L., C. W. Holden, and C. A. Trzcinka. 2017. What Are the Best Liquidity

Proxies for Global Research? Review of Finance 21:1355–1401.

Goldstein, M. A., and E. S. Hotchkiss. 2017. Proving Liquidity in an Illiquid Market: Dealer

Behavior in US Corporate Bonds. Boston College and Babson College Working Paper .

Goldstein, M. A., E. S. Hotchkiss, and E. R. Sirri. 2007. Transparency and Liquidity: A

Conrolled Experiment on Corporate Bonds. Review of Financial Studies 20:235–273.

Gorton, G. B., and G. G. Pennacchi. 1993. Security Baskets and Index-linked Securities.

The Journal of Business 66:1–27.

Grossman, S. J., and M. H. Miller. 1987. Liquidity and Market Structure. Journal of Finance

43:617–633.

Hamm, S. J. W. 2014. The Effect of ETFs on Stock Liquidity. the Ohio State University

Working paper .

Hendershott, T., and A. Madhavan. 2015. Click or Call? Auction versus Search in the

Over-the- Counter Market. Journal of Finance 70:419–447.

43

Holden, C. W. 1995. Index arbitrage as cross-sectional market making. Journal of Futures

Markets 15:423–455.

Holden, C. W., Y. Mao, and J. Nam. 2016. Price Discovery in the Stock and Corporate

Bond Markets. Cornell University and Indiana Unviersity Working Paper .

Hotchkiss, E., and T. Ronen. 2002. Informational efficiency of the corporate bond market:

an intraday analysis. Review of Financial Studies 15:1325–1354.

Huang, R., and H. Stoll. 1997. The components of the bid-ask spread: a general approach.

Review of Financial Studies 10:995–1034.

Huberman, G. 2001. Familiarity Breeds Investment. Review of Financial Studies 14:659–680.

Israeli, D., C. M. C. Lee, and S. Sridharan. 2017. Is there a Dark Side to Exchange Traded

Funds (ETF)? An Information Perspective. Review of Accounting Studies 22:1048–1083.

Jegadeesh, N., and A. Subrahmanyam. 1993. Liquidity Effects of the Introduction of the S&P

500 Index Futures Contract on the Underlying Stocks. Journal of Business 66:171–187.

Kang, J.-K., and R. A. M. Stulz. 1997. Why is There a Home Bias? An Analysis of Foreign

Portfolio Equity Ownership in Japan. Journal of Financial Economics 46:3–28.

Kyle, A. S. 1985. Continuous Auctions and Insider Trading. Econometrica 53:1315–1336.

Lesmond, D. A., J. Ogden, and C. A. Trzcinka. 1999. A New Estimate of Transaction Costs.

Review of Financial Studies 12:1113–1141.

Merton, R. C. 1987. A Simple Model of Capital Market Equilibrium with Incomplete Infor-

mation. The Journal of Finance 42:483–510.

Pan, K., and Y. Zeng. 2017. ETF Arbitrage under Liquidity Mismatch. Harvard University

and University of Washington Working Paper .

44

Pool, V. K., N. Stoffman, and S. E. Yonker. 2012. No Place Like Home: Familiarity in

Mutual Fund Manager Portfolio Choice. Review of Financial Studies 25:2563–2599.

Randall, O. 2015. Pricing and Liquidity in Over-The-Counter Markets. Emory University

Working Paper .

Rossi, M. 2014. Realized Volatility, Liquidity, and Corporate Yield Spreads. Quarterly

Journal of Finance 4.

Schestag, R., P. Schuster, and M. Uhrig-Homburg. 2016. Measuring Liquidity in Bond

Markets. Forthcoming in Review of Financial Studies .

Schultz, P. 2001. Corporate bond trading costs: a peek behind the curtain. Journal of

Finance 56:677–698.

Subrahmanyam, A. 1991. A Theory of Trading in Stock Index Futures. Review of Financial

Studies 4:17–51.

Wooldridge, J. M. 2010. Econometric Analysis of Cross Sectional and Panel Data. The MIT

Press.

Ye, S. 2019. How do ETFs Affect the Liquidity of the Underlying Corporate Bonds? Chinese

University of Hong Kong Working Paper .

45

Figure 1. Distribution of transaction costs and minimum trading quantities forindividual junk bonds. These bonds are held by the junk bond ETF portfolio (JNK) asof March 30, 2016. For each bond, the bid-ask spreads is divided by their mid point price.The average bid-ask spread is 110 basis points. The minimum trading quantities vary withbid or ask orders. The average minimum bid quantity $133,000, while the average minimumask quantity is $116,000. The median bid or ask quantity is $100,000. Data source: Fidelitycorporate bond database (“Depth of Book”).

46

Figure 2. Model overview for less accessible markets. (A) Before the introductionof ETFs, investors’ portfolios may have limited access to some market segments due to aminimum trading requirement, high transaction costs and other reasons. The cost of havinga diversified portfolio is high. These diversification costs increases as the portfolio is largeor the investors trade over multiple periods. The investors tend to have an under-diversifiedportfolio. (B) ETFs significantly lower the diversification cost than a portfolio of individualsecurities. First, investors flow into ETFs due to low cost of diversification. If the marketwas less accessible ex ante, then the influx of new investors to ETFs is greater. Second, ETFsare liquid. Third, large ETF liquidity is spilled over to the underlying securities througharbitrage mechanism. The liquidity of the underlying securities improves in the previouslyless accessible market.

47

Figure 3. Theoretical predictions after the introduction of ETFs. (A) If the marketwas highly accessible ex ante, discretionary uninformed investors can easily trade a portfolioof individual assets. With the introduction of ETFs, they prefer to trade ETFs in orderto reduce further the adverse selection cost. This migration of the uninformed investorscauses liquidity of underlying securities to decrease. (B) Suppose the market was partiallyaccessible, i.e. only security 1. Then the uninformed investors would prefer to migrate toETF, which will reduce the liquidity of security 1. At the same time, the uninformed investorswho hold the risk free assets would prefer to move investment from the risk free asset to ETF.The large ETF liquidity will spill over the underlying securities 1 and 2. Hence, the liquidityof security 1 is ambiguous, while the liquidity of security 2 improves. (C) If the market isinaccessible, then the liquidities of both security 1 and security 2 improve by applying similarlogic.

48

Figure 4. Newly-launched corporate bond ETFs and the underlying corporatebonds. The data sample is derived from the newly launched corporate bond ETFs and theunderlying corporate bonds over the period 2007 through 2014. The top graph shows theyearly aggregate number of corporate bond ETFs initiated, while the bottom graph showsthe number of the associated corporate bonds, whose trading were reported to TRACE. Theregular corporate bonds and 144A bonds are denoted by the blue and red colors, respec-tively. As of 2014Q4, 1253 distinct corporate bonds were traded and reported to TRACE.144A bonds that were only available to institutional investors through private placement areincreasingly included in ETFs since 2012.

(A)

(B)

49

Figure 5. The parallel trends for the treatment and control groups Each bondin the treatment group is matched with bonds in the control group. The bond liquidityexhibits the parallel trends in the pre-ETF period: (A) Parallel trends for non-investmentgrade bonds, (B) Parallel trends for investment grade bonds. In the post-ETF period, Theliquidity of non-investment grade bonds significantly improves as shown in (A).

(A)

(B)

50

Table

1.

Sum

mary

stati

stic

sof

fixed

inco

me

ET

Fs

Pan

elA

:B

lack

Rock

firs

tin

troduce

dfixed

inco

me

ET

Fs

in20

02,

follow

edby

Sta

teStr

eet,

Van

guar

d,

etc.

The

fund

fam

ily

den

oted

by

”Oth

ers”

incl

ude

AL

PS,C

laym

ore,

Col

um

bia

,D

BX

,E

GA

,F

lexShar

es,

Fra

nklin,

Gra

il,

Gugg

enhei

m,

Pow

erShar

es,

Pro

Shar

es,

and

Sch

wab

.T

he

num

ber

offixed

inco

me

ET

Fs

jum

ped

insi

nce

2007

.P

anel

B:

The

aver

age

num

ber

ofse

curi

ties

incl

uded

ina

fixed

inco

me

ET

Fva

ries

from

42to

137.

The

max

imum

is18

40.

The

num

ber

ofth

ese

curi

ties

firs

ten

tere

din

the

ET

Fp

ortf

olio

dra

mat

ical

lyin

crea

ses.

Pan

elA

.In

cepti

onof

ET

Fs

by

fund

fam

ily

Yea

rB

lack

Rock

Sta

teStr

eet

Van

guar

dO

ther

sW

isdom

Tre

eP

IMC

OA

dvis

orShar

esF

irst

Tru

stF

idel

ity

tota

l20

022

220

032

220

0716

45

530

2008

210

1325

2009

48

82

224

2010

213

11

118

2011

12

141

41

2320

126

315

12

2720

1310

22

85

11

332

2014

71

152

23

333

Tot

al50

2215

8023

125

63

216

Pan

elB

.E

TF

por

tfol

iosi

ze

Yea

rN

Mea

nM

inp10

Med

ian

p90

Max

Fir

sten

tere

d20

076

124

55

152

184

184

319

2009

375

11

9313

013

014

420

1011

429

1044

8191

452

2011

2084

126

4818

744

314

6420

1224

108

129

6522

267

620

3120

1329

175

131

105

417

1840

3804

2014

2413

721

2410

627

046

323

21

51

Table 2. Bond descriptive statistics. For credit rating, We assign a numeric score foreach credit rating by each of the three rating agencies – S& P, Moody’s, Fitch – followingBecker and Ivashina (2015). At each given time, we merge the three credit rating scoresas follows: If the three ratings are available, we take the middle rating. If two ratingsare available, we take the lower. Non-ETF bonds are matched with ETF bonds using issueamount, credit rating, maturity at issuance, age at the inception date of ETF, and cumulativetrading volume during 6 months prior to the inception date of ETF.

ETF bonds non-ETF bonds DiD p-valueDiD=0

Total number of bond issues 2,623 2,623Average issue size ($M) 1,007 1,032 −25 0.23≤ 100M 429 (16%) 517 (20%)100-500M 726 (28%) 692 (26%)> 500M 1,468 (56%) 1,428 (54%)

Average credit quality 19.37 19.20 0.17 0.33AA− and up 283 (11%) 351 (13%)BBB− thru A+ 1,729 (66%) 1,554 (59%)below BBB− 625 (24%) 732 (28%)

Average maturity at issue 7.86 7.71 0.15 0.13(in years)< 2 years 21 (1%) 105 (4%)2-5 years 465 (18%) 597 (23%)5-20 years 2,048 (78%) 1,845 (70%)> 20 years 103 (4%) 90 (3%)

Age at ETF inception date 1.50 1.52 −0.02 0.20(in years)< 2 years 1,955 (74%) 1,959 (74%)2-5 years 663 (25%) 617 (23%)≥ 5 years 19 (1%) 61 (2%)

52

Table 3. Bond trading characteristics. Transaction price, transaction volume, trans-action frequency. To compute average transaction cost per day, we first use a subset ofcorporate bonds whose buy and sell transactions are observed for each day. We computethe average sell price subtracted by the average buy price. To estimate the average tradingcost per transaction, we use the last trading price on the most recent transaction day asa benchmark price and examine the changes of trading prices on the current day relativeto the benchmark price, following the insight of Huang and Stoll (1997); Hendershott andMadhavan (2015). Transaction costs are further examined in details based on the cumulativetrading volume during 6 months prior to the inception date of ETF.

ETF bonds non-ETF bondsbefore after before after

Average trade price(% of par value) 101.11 101.01 100.40 99.93Total trading volume ($B) 1341.73 1122.89 1009.62 788.94Total number of trades (1000) 2304.62 2364.67 2143.66 1913.70< $100K 1291.74 1419.42 1280.89 1194.49$100K-1M 547.10 565.50 466.48 403.93$1-5M 351.69 291.44 321.79 261.44>$5M 114.10 88.31 74.50 53.84

Average two-way transaction cost per day (bp) 65.72 67.21 75.42 76.74Average transaction cost per transaction (bp)

VOL 1 (small) 47.69 32.78 42.75 52.44VOL 2 27.11 22.54 27.64 30.85VOL 3 21.82 22.52 25.08 24.20VOL 4 20.18 20.00 24.15 28.78VOL 5 (large) 20.15 21.45 17.81 19.72

53

Table 4. Full sample analysis. The effective half spreads of ETF-bonds are comparedwith those of non-ETF bonds. Following the methodology of Bessembinder et al. (2006),the first stage is estimated as Qt = a + bQt−1 + εt, where Qt is +1 for the buy order, −1for the sell order, and 0 for the inter-dealer orders. We write the unexpected order flowεt as Q∗

t . The second stage is estimated as ∆Pt = a + γSQ∗t + αS∆Qt + wXt + ωt. To

assess the changes of effective spreads of ETF-bonds relative to the changes of effectivespreads of non-ETF bonds, we interact ∆Qt with Post and ETF . The non-informationalcomponent αS is divided into the changes of the effective spreads in the treatment groupand the changes of the effective spreads in the control group: α1∆Qt + α2S∆Qt × Post +α3S∆Qt × ETF + α4S∆Qt × Post × ETF. Of our main focus is the coefficient α4S. Weinclude three public information variables, each of which is measured on the date of the mostrecent transaction on a day prior to the current transaction. The changes in the interestrate for on-the-run Treasury security is matched to the corporate bond based on maturity.The percentage return on the issuing firm’s common stock. The change in the default spreadbetween long-term indexes of BAA-rated bonds and U. S. Treasury securities. we interactthese variables with investment- and non-investment-grade indicator variables to account forpotential differences in sensitivity based on the bond’s risk. We estimate the first and secondstage regressions using the weighted least squares, where the weight is inversely related tothe number of days elapsed between Qt and Qt−1. The statistical significance is measuredthe probability from block bootstrapping, following Bessembinder et al. (2006).

coeff p-valueFirst stage regressionIntercept 0.1137∗∗∗ 0.000Qt−1 0.0678∗∗∗ 0.000

Second stage regressionIntercept 0.0276∗ 0.060Q∗ 0.0316∗∗∗ 0.000Treasury return −0.0119∗∗∗ 0.000Stock return × invgrd 0.0176∗∗∗ 0.000Stock return × noninvgrd 0.0745∗∗∗ 0.000Default spread × invgrd −0.0074∗∗∗ 0.000Default spread × noninvgrd 0.0386∗∗∗ 0.000∆Q 0.4221∗∗∗ 0.000∆Q× Post 0.0262∗∗∗ 0.000∆Q× ETF −0.0537∗∗∗ 0.000∆Q× Post × ETF −0.0132∗∗∗ 0.000

R2 0.1804RMSE 0.7324N of block boostrapping 50

54

Table 5. By Arbitrage activities To estimate the arbitrage activities of each ETF,we add the magnitude of the daily changes of the ETF shares outstanding in the postevent period, relative to the average ETF shares outstanding. If the arbitrage activitiesfall below (or above) the 40th percentile, i.e. 125%, we consider it low (or high) arbitrageETFs. The effective half spreads of ETF-bonds are compared with those of non-ETF bonds.Following the methodology of Bessembinder et al. (2006), the first stage is estimated asQt = a + bQt−1 + εt, where Qt is +1 for the buy order, −1 for the sell order, and 0 forthe inter-dealer orders. We write the unexpected order flow εt as Q∗

t . The second stage isestimated as ∆Pt = a+γSQ∗

t +αS∆Qt+wXt+ωt. To assess the changes of effective spreadsof ETF-bonds relative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt

with Post and ETF . The non-informational component αS is divided into the changes ofthe effective spreads in the treatment group and the changes of the effective spreads inthe control group: α1∆Qt + α2S∆Qt × Post + α3S∆Qt × ETF + α4S∆Qt × Post× ETF.Of our main focus is the coefficient α4S. We include three public information variablesXt as discussed in 3.5. The statistical significance is measured the probability from blockbootstrapping, following Bessembinder et al. (2006).

coeff p-valueFirst stage regressionIntercept 0.1074∗∗∗ 0.000Qt−1 0.0685∗∗∗ 0.000

Second stage regressionIntercept 0.0372∗∗∗ 0.000Q∗ 0.0325∗∗∗ 0.000Treasury return −0.0157∗∗∗ 0.000Stock return × invgrd 0.0262∗∗∗ 0.000Stock return × noninvgrd 0.0780∗∗∗ 0.000Default spread × invgrd −0.0086∗∗∗ 0.000Default spread × noninvgrd 0.0478∗∗∗ 0.000Arb H × ∆Q 0.4266∗∗∗ 0.000Arb H × ∆Q× Post 0.0227∗∗∗ 0.000Arb H × ∆Q× ETF 0.0227∗∗∗ 0.000Arb H × ∆Q× Post × ETF −0.0300∗∗∗ 0.000Arb L × ∆Q 0.4143∗∗∗ 0.000Arb L × ∆Q× Post 0.0210∗∗∗ 0.000Arb L × ∆Q× ETF −0.0486∗∗∗ 0.000Arb L × ∆Q× Post × ETF 0.0119∗∗∗ 0.000

R2 0.1874N of block bootstrapping 50

55

Table

6.

By

Arb

itra

ge

and

Ord

er

Imbala

nce

inti

me

seri

es

To

esti

mat

eth

ear

bit

rage

opp

ortu

nit

ies

ofea

chE

TF

,w

ees

tim

ate

adiff

eren

ceb

etw

een

the

ET

Fm

arke

tpri

cep

ersh

are

and

its

net

asse

tva

lue

(NA

V)

per

shar

eon

the

pre

vio

us

day

,w

hic

his

then

div

ided

by

the

NA

V.

Ifth

ear

bit

rage

opp

ortu

nit

ies

fall

bel

ow(o

rab

ove)

the

25th

per

centi

le,

we

consi

der

itlo

w(o

rhig

h)

arbit

rage

ET

Fs.

To

esti

mat

eor

der

imbal

ance

,w

eco

nst

ruct

dol

lar

buy

and

sell

volu

mes

per

day

ofal

lth

eco

nst

ituen

tb

onds

ofea

chE

TF

.If

the

diff

eren

ceof

buy

and

sell

dol

lar

volu

mes

isb

elow

(or

abov

e)th

e25

thp

erce

nti

le,

we

consi

der

the

order

imbal

ance

low

(or

hig

h).

Fol

low

ing

the

met

hodol

ogy

ofB

esse

mbin

der

etal

.(2

006)

,th

efirs

tst

age

ises

tim

ated

asQt

=a

+bQ

t−1

+ε t,

wher

eQt

is+

1fo

rth

ebuy

order

,−

1fo

rth

ese

llor

der

,an

d0

for

the

inte

r-dea

ler

order

s.W

ew

rite

the

unex

pec

ted

order

flow

ε tas

Q∗ t.

The

seco

nd

stag

eis

esti

mat

edas

∆Pt

=a

+γSQ

∗ t+αS

∆Qt+wXt+ωt.

The

non

-info

rmat

ional

com

pon

entαS

isdiv

ided

into

the

chan

ges

ofth

eeff

ecti

vesp

read

sin

the

trea

tmen

tgr

oup

and

the

chan

ges

ofth

eeff

ecti

vesp

read

sin

the

contr

olgr

oup:α1S

∆Qt+α2S

∆Qt×ETF.

Of

our

mai

nfo

cus

isth

eco

effici

entα2S

.

Arb

itra

geO

pp

ortu

nit

yH

igh

Hig

hL

owL

owO

rder

Imbal

ance

Hig

hL

owH

igh

Low

coeff

p-va

lue

coeff

p-va

lue

coeff

p-va

lue

coeff

p-va

lue

Fir

stst

age

regr

essi

onIn

terc

ept

0.09

79∗∗

∗0.

000

0.10

13∗∗

∗0.

000

0.09

27∗∗

∗0.

000

0.09

73∗∗

∗0.

000

Qt−

10.

0598

∗∗∗

0.00

00.

0616

∗∗∗

0.00

00.

0643

∗∗∗

0.00

00.

0636

∗∗∗

0.00

0

Sec

ond

stag

ere

gres

sion

Q∗

0.02

60∗∗

∗0.

002

0.01

54∗∗

∗0.

000

0.01

99∗∗

∗0.

002

0.02

59∗∗

∗0.

000

Tre

asury

retu

rn−

0.01

16∗∗

∗0.

000−

0.00

80∗∗

∗0.

000−

0.02

66∗∗

∗0.

000−

0.03

02∗∗

∗0.

000

Sto

ckre

turn×

invgr

d0.

0159

∗∗∗

0.00

00.

0177

∗∗∗

0.00

00.

0157

∗∗∗

0.00

00.

0124

∗∗∗

0.00

0Sto

ckre

turn×

non

invgr

d0.

0584

∗∗∗

0.00

00.

0507

∗∗∗

0.00

00.

0657

∗∗∗

0.00

00.

0732

∗∗∗

0.00

0D

efau

ltsp

read×

invgr

d−

0.01

70∗∗

∗0.

000−

0.01

10∗∗

∗0.

000−

0.01

58∗∗

∗0.

000−

0.00

19∗∗

∗0.

000

Def

ault

spre

ad×

non

invgr

d0.

0126

∗∗∗

0.00

00.

0510

∗∗∗

0.00

00.

0589

∗∗∗

0.00

00.

0308

∗∗∗

0.00

0∆Q

0.42

93∗∗

∗0.

000

0.44

53∗∗

∗0.

000

0.43

22∗∗

∗0.

000

0.42

95∗∗

∗0.

000

∆Q×

ET

F−

0.07

44∗∗

∗0.

000−

0.06

44∗∗

∗0.

000−

0.06

54∗∗

∗0.

000−

0.05

56∗∗

∗0.

000

R2

0.17

880.

1872

0.18

200.

1868

Nof

blo

ckb

oot

stra

ppin

g50

050

050

050

0

56

Table 7. Changes of the effective half spreads by cumulative trading volume.The bonds are segmented into the quintiles based on their cumulative trading volume during6 months prior to ETFs. Following the methodology of Bessembinder et al. (2006), the firststage is estimated as Qt = a + bQt−1 + εt, where Qt is +1 for the buy order, −1 for thesell order, and 0 for the inter-dealer orders. We write the unexpected order flow εt as Q∗

t .The second stage is estimated as ∆Pt = a + γSQ∗

t + αS∆Qt + wXt + ωt. To assess thechanges of effective spreads of ETF-bonds relative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt with Post and ETF . The non-informational component αSis divided into the changes of the effective spreads in the treatment group and the changesof the effective spreads in the control group: α1∆Qt + α2S∆Qt× Post+ α3S∆Qt×ETF +α4S∆Qt×Post×ETF. Of our main focus is the coefficient α4S. We include the three publicinformation variables Xt as discussed in section 3.5. The statistical significance is measuredthe probability from block bootstrapping, following Bessembinder et al. (2006).

57

coeff p-valueFirst stage regressionIntercept 0.1137∗∗∗ 0.000Qt−1 0.0678∗∗∗ 0.000

Second stage regressionIntercept 0.0268∗∗∗ 0.000Q∗ 0.0316∗∗∗ 0.000Treasury return −0.0116∗∗∗ 0.000Stock return × invgrd 0.0175∗∗∗ 0.000Stock return × noninvgrd 0.0745∗∗∗ 0.000Default spread × invgrd −0.0075∗∗∗ 0.000Default spread × noninvgrd 0.0389∗∗ 0.040VOL 1 (Small) ×∆Q 0.9263∗∗∗ 0.000VOL 1 ×∆Q× Post −0.0009 0.300VOL 1 ×∆Q× ETF −0.4695∗∗∗ 0.000VOL 1 ×∆Q× Post × ETF −0.0881∗∗∗ 0.000VOL 2 ×∆Q 0.8714∗∗∗ 0.000VOL 2 ×∆Q× Post 0.0141∗∗ 0.020VOL 2 ×∆Q× ETF −0.5017∗∗∗ 0.000VOL 2 ×∆Q× Post × ETF −0.0355∗∗∗ 0.000VOL 3 ×∆Q 0.7360∗∗∗ 0.000VOL 3 ×∆Q× Post −0.0384∗∗∗ 0.000VOL 3 ×∆Q× ETF −0.3180∗∗∗ 0.000VOL 3 ×∆Q× Post × ETF 0.0185∗∗∗ 0.000VOL 4 ×∆Q 0.4440∗∗∗ 0.000VOL 4 ×∆Q× Post 0.0394∗∗∗ 0.000VOL 4 ×∆Q× ETF −0.1036∗∗∗ 0.000VOL 4 ×∆Q× Post × ETF −0.0015 0.240VOL 5 (Large) ×∆Q 0.4159∗∗∗ 0.000VOL 5 ×∆Q× Post 0.0237∗∗∗ 0.000VOL 5 ×∆Q× ETF −0.0528∗∗∗ 0.000VOL 5 ×∆Q× Post × ETF −0.0011 0.180

R2 0.1822RMSE 0.7316N of block bootstrapping 100

58

Table 8. Changes of the effective half spreads by credit rating of the bonds. Thebonds are divided into two groups based on the bond credit ratings. We assign numericscores to the bond credit ratings from S&P, Moody’s, and Fitch, following Bo and Ivashina(2012). For each bond, if two credit ratings are available, we take the lower. If threecredit ratings are available, we take the middle. Investment grades bonds have credit ratingsBBB− or above based on S&P or Fitch (or equivalently Baa3 from Moody’s). Following themethodology of Bessembinder et al. (2006), the first stage is estimated as Qt = a+bQt−1+εt,where Qt is +1 for the buy order, −1 for the sell order, and 0 for the inter-dealer orders.We write the unexpected order flow εt as Q∗

t . The second stage is estimated as ∆Pt =a + γSQ∗

t + αS∆Qt + wXt + ωt. To assess the changes of effective spreads of ETF-bondsrelative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt with Postand ETF . The non-informational component αS is divided into the changes of the effectivespreads in the treatment group and the changes of the effective spreads in the control group:α1∆Qt + α2S∆Qt × Post + α3S∆Qt × ETF + α4S∆Qt × Post× ETF. Of our main focusis the coefficient α4S. We include the three public information variables Xt as discussed insection 3.5. The statistical significance is measured the probability from block bootstrapping,following Bessembinder et al. (2006).

Investment grade Non-investment gradecoeff p-value coeff p-value

First stage regressionIntercept 0.1190∗∗∗ 0.000 0.1003∗∗∗ 0.000Qt−1 0.0672∗∗∗ 0.000 0.0684∗∗∗ 0.000

Second stage regressionIntercept 0.0072∗∗∗ 0.000 0.1053∗∗∗ 0.000Q∗ 0.0172∗∗∗ 0.000 0.0842∗∗∗ 0.000Treasury return −0.0101∗∗∗ 0.000 −0.0270∗∗∗ 0.000Stock return × invgrd 0.0178∗∗∗

Stock return × noninvgrd 0.0747∗∗∗ 0.000Default spread × invgrd 0.0067∗∗∗

Default spread × noninvgrd 0.0341∗∗∗ 0.000∆Q 0.3728∗∗∗ 0.000 0.5695∗∗∗ 0.000∆Q× Post 0.0194∗∗∗ 0.000 0.0214∗∗∗ 0.000∆Q× ETF −0.0347∗∗∗ 0.000 −0.0200∗∗∗ 0.000∆Q× Post × ETF 0.0001 0.200 −0.0542∗∗∗ 0.000

R2 0.1603 0.2291RMSE 0.5493 0.9049N of block bootstrapping 100 100

59

Table 9. Changes of the effective half spreads of bonds by the time to maturity.The time to maturity is measured as the number of days to the maturity on the first offerdate of ETF. The bonds are segmented into two groups with a threshold of 2 years i.e. 730days. Following the methodology of Bessembinder et al. (2006), the first stage is estimatedas Qt = a + bQt−1 + εt, where Qt is +1 for the buy order, −1 for the sell order, and 0 forthe inter-dealer orders. We write the unexpected order flow εt as Q∗

t . The second stage isestimated as ∆Pt = a+γSQ∗

t +αS∆Qt+wXt+ωt. To assess the changes of effective spreadsof ETF-bonds relative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt

with Post and ETF . The non-informational component αS is divided into the changes ofthe effective spreads in the treatment group and the changes of the effective spreads in thecontrol group: α1∆Qt+α2S∆Qt×Post+α3S∆Qt×ETF +α4S∆Qt×Post×ETF. Of ourmain focus is the coefficient α4S. We include the three public information variables Xt asdiscussed in section 3.5. The statistical significance is measured the probability from blockbootstrapping, following Bessembinder et al. (2006).

Life remaining ≤ 2 years Life remaining ≥ 2 yearscoeff p-value coeff p-value

First stage regressionIntercept −0.0529∗∗∗ 0.000 0.1298∗∗∗ 0.000Qt−1 0.0354∗∗∗ 0.000 0.0667∗∗∗ 0.000

Second stage regressionIntercept −0.0043 0.290 0.0354∗∗∗ 0.000Q∗ 0.0006∗∗∗ 0.000 0.0385∗∗∗ 0.000Treasury return 0.0160∗∗∗ 0.000 −0.0144∗∗∗ 0.000Stock return × invgrd 0.0041∗∗∗ 0.000 0.0194∗∗∗ 0.000Stock return × noninvgrd 0.0713∗∗∗ 0.000 0.0744∗∗∗ 0.000Default spread × invgrd −0.0028∗∗∗ 0.000 −0.0084∗∗∗ 0.000Default spread × noninvgrd −0.0661∗∗∗ 0.000 0.0393∗∗∗ 0.000∆Q 0.1591∗∗ 0.040 0.4547∗∗∗ 0.000∆Q× Post −0.0246∗∗∗ 0.000 0.0353∗∗∗ 0.000∆Q× ETF −0.0539∗∗∗ 0.000 −0.0861∗∗∗ 0.000∆Q× Post × ETF 0.0989∗∗∗ 0.000 −0.0190∗∗∗ 0.000

R2 0.0762 0.1936RMSE 0.3026 0.7270N of block bootstrapping 100 100

60

Table 10. Quasi-natural experiment for HYG. We investigate changes of the liquidityof the constituent bonds for HYG around the index rule change during 2009. The effectivehalf spreads of ETF-bonds are compared with those of non-ETF bonds. Following themethodology of Bessembinder et al. (2006), the first stage is estimated as Qt = a+bQt−1+εt,where Qt is +1 for the buy order, −1 for the sell order, and 0 for the inter-dealer orders.We write the unexpected order flow εt as Q∗

t . The second stage is estimated as ∆Pt =a + γSQ∗

t + αS∆Qt + wXt + ωt.To assess the changes of effective spreads of ETF-bondsrelative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt with Postand ETF . The non-informational component αS is divided into the changes of the effectivespreads in the treatment group and the changes of the effective spreads in the control group:α1∆Qt + α2S∆Qt × Post + α3S∆Qt × ETF + α4S∆Qt × Post× ETF. Of our main focusis the coefficient α4S. We include the three public information variables Xt as discussed insection 3.5. The statistical significance is measured the probability from block bootstrapping,following Bessembinder et al. (2006).

coeff p-valueFirst stage regressionIntercept 0.0485∗∗∗ 0.00Qt−1 0.0686∗∗∗ 0.00

Second stage regressionIntercept 0.1090∗∗∗ 0.00Q∗ 0.0964∗∗∗ 0.00Treasury return −0.0052∗∗∗ 0.00Stock return × invgrd 0.0054∗∗∗ 0.00Stock return × noninvgrd 0.0025∗∗∗ 0.00Default spread × invgrd 0.0195∗∗∗ 0.00Default spread × noninvgrd 0.1943∗∗∗ 0.00∆Q 0.4318∗∗∗ 0.00∆Q× Post 0.0686∗∗∗ 0.00∆Q× ETF 0.2083∗∗∗ 0.00∆Q× Post × ETF −0.0150∗∗∗ 0.00

R2 0.1469RMSE 0.9876N of block boostrapping 100

61

Table 11. ETF by management styles. ETFs can be actively or passively managed. Weseparate the bonds included by active or index ETFs. The bonds can be owned by both stylesof the ETFs, which we exclude from the analysis. The effective half spreads of ETF-bondsare compared with those of non-ETF bonds. Following the methodology of Bessembinderet al. (2006), the first stage is estimated as Qt = a + bQt−1 + εt, where Qt is +1 for thebuy order, −1 for the sell order, and 0 for the inter-dealer orders. We write the unexpectedorder flow εt as Q∗

t . The second stage is estimated as ∆Pt = a+ γSQ∗t +αS∆Qt +wXt +ωt.

To assess the changes of effective spreads of ETF-bonds relative to the changes of effectivespreads of non-ETF bonds, we interact ∆Qt with Post and ETF . The non-informationalcomponent αS is divided into the changes of the effective spreads in the treatment groupand the changes of the effective spreads in the control group: α1∆Qt + α2S∆Qt × Post +α3S∆Qt × ETF + α4S∆Qt × Post × ETF. Of our main focus is the coefficient α4S. Weinclude the three public information variables Xt as discussed in section 3.5. The statisticalsignificance is measured the probability from block bootstrapping, following Bessembinderet al. (2006).

Active ETF Index ETFcoeff p-value coeff p-value

First stage regressionIntercept 0.1490∗∗∗ 0.000 0.0857∗∗∗ 0.000Qt−1 0.0640∗∗∗ 0.000 0.0473∗∗∗ 0.000

Second stage regressionIntercept 0.0092∗∗∗ 0.000 −0.0169∗∗∗ 0.000Q∗ 0.0143∗∗∗ 0.000 0.0518∗∗∗ 0.000Tresury return −0.0084∗∗∗ 0.000 0.0240∗∗∗ 0.000Stock return × Invgrd 0.0150∗∗∗ 0.000 0.0350∗∗∗ 0.000Stock return × Noninvgrd 0.0500∗∗∗ 0.000 0.0890∗∗∗ 0.000Default spread × Invgrd −0.0124∗∗∗ 0.000 0.0024∗∗∗ 0.000Default spread × Noninvgrd 0.0271∗∗∗ 0.000 0.0490∗∗∗ 0.000∆Q 0.4310∗∗∗ 0.000 0.3510∗∗∗ 0.000∆Q× Post 0.0086∗∗∗ 0.000 0.0639∗∗∗ 0.000∆Q× ETF −0.0530∗∗∗ 0.000 0.0083∗∗∗ 0.000∆Q× Post× ETF 0.0080∗∗∗ 0.000 −0.1028∗∗∗ 0.000

R2 0.1813 0.1851RMSE 0.6364 0.6762N of block bootstrapping 100 100

62

Table 12. Multivariate analysis. We evaluate the relative strength of the explanatoryvariables – cumulative trading volume, bond credit rating, age, ETF styles – to the liquidityof the underlying bonds. The effective half spreads of ETF-bonds are compared with those ofnon-ETF bonds. Following the methodology of Bessembinder et al. (2006), the first stage isestimated as Qt = a+bQt−1+εt, where Qt is +1 for the buy order, −1 for the sell order, and 0for the inter-dealer orders. We write the unexpected order flow εt as Q∗

t . The second stage isestimated as ∆Pt = a+γSQ∗

t +αS∆Qt+wXt+ωt. To assess the changes of effective spreadsof ETF-bonds relative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt

with Post and ETF . The non-informational component αS is divided into the changes ofthe effective spreads in the treatment group and the changes of the effective spreads in thecontrol group: α1∆Qt+α2S∆Qt×Post+α3S∆Qt×ETF +α4S∆Qt×Post×ETF. Of ourmain focus is the coefficient α4S. We include the three public information variables Xt asdiscussed in section 3.5. The statistical significance is measured the probability from blockbootstrapping, following Bessembinder et al. (2006).

63

All Invgrad Non-invgrdcoeff p-value coeff p-value coeff p-value

First stage regressionIntercept 0.1137∗∗∗ 0.000 0.1190∗∗∗ 0.000 0.1003∗∗∗ 0.000Qt−1 0.0678∗∗∗ 0.000 0.0672∗∗∗ 0.000 0.0684∗∗∗ 0.000

Second stage regressionIntercept 0.0225∗∗∗ 0.000 0.0024∗∗∗ 0.000 0.1034∗∗∗ 0.000Q∗ 0.0360∗∗∗ 0.000 0.0213∗∗∗ 0.000 0.0858∗∗∗ 0.000Treasury return −0.0088∗∗∗ 0.000 −0.0070∗∗∗ 0.000 −0.0259∗∗∗ 0.000Stock return × Invgrd 0.0176∗∗∗ 0.000 0.0178∗∗∗ 0.000Stock return × Noninvgrd 0.0742∗∗∗ 0.000 0.0746∗∗∗ 0.000Default spread × Invgrd −0.0079∗∗∗ 0.000 −0.0073∗∗∗ 0.000Default spread × Noninvgrd 0.0383∗∗∗ 0.000 0.0336∗∗∗ 0.000VOL 1 (Small) ×∆Q 0.9692∗∗∗ 0.000 0.9814∗∗∗ 0.000 0.6823∗∗∗ 0.000VOL 1 ×∆Q× Post 0.0056∗∗∗ 0.000 0.0087∗∗∗ 0.000 0.0604∗∗∗ 0.000VOL 1 ×∆Q× ETF −0.5092∗∗∗ 0.000 −0.5623∗∗∗ 0.000 −0.1966∗∗∗ 0.000VOL 1 ×∆Q× Post× ETF −0.0902∗∗∗ 0.000 −0.0482∗∗∗ 0.000 −0.1979∗∗∗ 0.000VOL 3 ×∆Q 0.7490∗∗∗ 0.000 0.9482∗∗∗ 0.000 0.3548∗∗∗ 0.000VOL 3 ×∆Q× Post −0.0271∗∗∗ 0.000 0.0028∗∗∗ 0.000 0.0070∗∗∗ 0.000VOL 3 ×∆Q× ETF −0.3178∗∗∗ 0.000 −0.6446∗∗∗ 0.000 −0.3601∗∗∗ 0.000VOL 3 ×∆Q× Post× ETF −0.0150∗∗∗ 0.000 0.0266∗∗∗ 0.000 −0.1078∗∗∗ 0.000VOL 5 (Large) ×∆Q 0.4504∗∗∗ 0.000 0.4019∗∗∗ 0.000 0.5844∗∗∗ 0.000VOL 5 ×∆Q× Post 0.0333∗∗∗ 0.000 0.0276∗∗∗ 0.000 0.0289∗∗∗ 0.000VOL 5 ×∆Q× ETF −0.0895∗∗∗ 0.000 −0.0386∗∗∗ 0.000 −0.0803∗∗∗ 0.000VOL 5 ×∆Q× Post× ETF 0.0091 0.000 −0.0039 0.000 0.2300∗∗∗ 0.000Life ≤ 2yr ×∆Q −0.3139∗∗∗ 0.000 −0.2789∗∗∗ 0.000 −0.0421∗∗∗ 0.000Life ≤ 2yr ×∆Q× Post −0.0589∗∗∗ 0.000 −0.0548∗∗∗ 0.000 −0.1876∗∗∗ 0.000Life ≤ 2yr ×∆Q× ETF 0.0062∗∗∗ 0.000 0.0326∗∗∗ 0.000 0.2370∗∗∗ 0.000Life ≤ 2yr ×∆Q× Post× ETF 0.1351∗∗∗ 0.000 0.0839∗∗∗ 0.000 0.3524∗∗∗ 0.000

R2 0.1910 0.1755 0.2310RMSE 0.6729 0.7774 1.0202N of block bootstrapping 100 100 100

64

Table

13.

144A

vs.

publi

cb

onds

tradin

gch

ara

cteri

stic

s.14

4Aan

dpublic

bon

ds

com

par

ison

ofav

erag

edai

lytr

adin

gvo

lum

ein

million

dol

lars

and

issu

eam

ount

inbillion

dol

lars

.In

vest

men

tgr

ade

bon

ds

hav

ecr

edit

rati

ngsBBB−

and

abov

e.T

he

aver

age

dai

lytr

adin

gvo

lum

eex

plo

ded

duri

ng

the

3rd

quar

ter

of20

14.

This

was

pre

vale

nt

bot

hin

inve

stm

ent

grad

ean

dnon

-inve

stm

ent

grad

eb

onds.

Ave

rage

dai

lytr

adin

gvo

lum

eIs

sue

Am

ounts

Full

Inv

grad

eN

onin

vgr

dIn

vgr

dN

onin

vgr

dY

ear

Qtr

.14

4AP

ublic

144A

Public

144A

Public

144A

Public

144A

Public

2008

41

9262

0%0

01

3386

0%10

9216

636

8%11

6863

7318

%20

104

811

742

0%20

7467

0%2

4274

0%13

5823

941

7%13

7162

8622

%20

114

311

455

0%3

7137

0%0

4318

676

1857

35%

1021

5371

19%

2012

41

9483

0%0

6042

140

880%

597

5304

8%10

3811

633

9%20

134

810

417

0%4

4993

0%8

5611

0%18

812

6915

%65

848

0714

%20

141

412

306

0%4

5360

0%4

7153

0%20

142

5711

339

1%29

4338

1%53

7250

1%20

143

2440

1063

823

%36

637

7110

%21

0971

8429

%20

144

2558

1181

722

%35

240

679%

2261

7953

28%

2015

128

4613

481

21%

351

4549

8%25

3193

3027

%20

152

2896

1285

523

%29

237

798%

2675

9219

29%

2015

327

3511

494

24%

253

3135

8%25

2486

1729

%20

154

2765

1109

625

%23

928

928%

2561

8377

31%

65

Table

14.

Liq

uid

ity

of

144A

bonds.

Hal

f-sp

read

sby

all,

volu

me

and

cred

itra

ting

wit

hth

e14

4Ab

onds.

The

firs

tst

age

ises

tim

ated

asQt

=a

+bQ

t−1

+ε t,

wher

eQt

is+

1fo

rth

ebuy

order

,−

1fo

rth

ese

llor

der

,an

d0

for

the

inte

r-dea

ler

order

s.W

ew

rite

the

unex

pec

ted

order

flow

ε tas

Q∗ t.

The

seco

nd

stag

eis

esti

mat

edas

∆Pt

=a

+γSQ

∗ t+αS

∆Qt+wXt+ωt.

To

asse

ssth

ech

ange

sof

effec

tive

spre

ads

ofE

TF

-144

Ab

onds

rela

tive

toth

ech

ange

sof

effec

tive

spre

ads

ofnon

-ET

F14

4Ab

onds,

we

inte

ract

∆Qt

wit

hPost

andETF

.T

he

non

-info

rmat

ional

com

pon

ent

αS

isdiv

ided

into

the

chan

ges

ofth

eeff

ecti

vesp

read

sin

the

trea

tmen

tgr

oup

and

the

chan

ges

ofth

eeff

ecti

vesp

read

sin

the

contr

olgr

oup:α1∆Qt

+α2S

∆Qt×Post

+α3S

∆Qt×ETF

+α4S

∆Qt×Post×ETF.

Of

our

mai

nfo

cus

isth

eco

effici

entα4S

.W

ein

clude

the

thre

epublic

info

rmat

ion

vari

able

sXt

asdis

cuss

edin

sect

ion

3.5.

The

stat

isti

cal

sign

ifica

nce

ism

easu

red

the

pro

bab

ilit

yfr

omblo

ckb

oot

stra

ppin

g,fo

llow

ing

Bes

sem

bin

der

etal

.(2

006)

.

66

Full

By

volu

me

By

cred

itra

ting

Hig

h/L

owV

olIn

vest

men

tgr

ade

Non

-inve

stm

ent

grad

eco

effp-

valu

eco

effp-

valu

eco

effp-

valu

eco

effp-

valu

eF

irst

stag

ere

gres

sion

Inte

rcep

t−

0.03

89∗∗

∗0.

000−

0.03

89∗∗

∗0.

000−

0.06

70∗∗

0.00

0−

0.01

68∗

0.09

0Qt−

10.

0299

∗∗0.

010

0.02

99∗∗

0.01

00.

0181

∗∗0.

100

0.03

57∗∗

0.01

0

Sec

ond

stag

ere

gres

sion

Inte

rcep

t−

0.04

08∗∗

0.05

0−

0.04

77∗∗

0.03

0−

0.09

74∗∗

∗0.

000−

0.01

480.

036

Q∗

0.10

04∗∗

0.00

00.

0976

∗∗∗

0.00

00.

0247

∗0.

060

0.11

15∗∗

∗0.

000

Tre

asury

retu

rn−

0.00

050.

540

0.00

450.

310

0.05

84∗∗

∗0.

000−

0.01

190.

430

Sto

ckre

turn×

invgr

d0.

0025

0.48

0−

0.00

59∗∗

0.02

0−

0.00

39∗∗

0.02

0Sto

ckre

turn×

non

invgr

d0.

0597

∗∗∗

0.00

00.

0572

∗∗∗

0.00

00.

0600

∗∗∗

0.00

0D

efau

ltsp

read×

invgr

d−

1.55

79∗∗

∗0.

000−

1.35

69∗∗

∗0.

000−

1.23

86∗∗

∗0.

000

Def

ault

spre

ad×

non

invgr

d0.

0675

0.13

00.

0032

0.14

00.

0881

0.23

0∆Q

0.02

31∗

0.07

00.

0321

∗∗∗

0.00

00.

0202

0.11

0∆Q×

Pos

t0.

0267

0.10

00.

0306

∗∗∗

0.00

00.

0266

0.14

0∆Q×

ET

F0.

0716

∗∗0.

000

0.00

520.

230

0.12

97∗∗

∗0.

000

∆Q×

Pos

t×

ET

F−

0.04

92∗∗

0.02

0−

0.03

51∗∗

0.01

0−

0.09

71∗∗

∗0.

000

VO

LL×

∆Q

0.18

28∗∗

∗0.

000

VO

LL×

∆Q×

Pos

t−

0.01

140.

360

VO

LL×

∆Q×

ET

F−

0.12

91∗∗

∗0.

000

VO

LL×

∆Q×

Pos

t×

ET

F0.

0879

∗∗0.

010

VO

LH×

∆Q

−0.

0277

0.11

0V

OL

H×

∆Q×

Pos

t0.

0401

∗∗0.

010

VO

LH×

∆Q×

ET

F0.

1349

∗∗∗

0.00

0V

OL

H×

∆Q×

Pos

t×

ET

F−

0.12

24∗∗

∗0.

000

R2

0.11

910.

1406

0.21

860.

1245

Nof

blo

ckb

oos

trap

pin

g10

010

010

010

0

67

Table 15. Control for firm fixed effect. To control for unobservable bond characteristicsrelated to the bond issuing firms, we include firm fixed effects in the analysis. The effectivehalf spreads of ETF-bonds are compared with those of non-ETF bonds. Following themethodology of Bessembinder et al. (2006), the first stage is estimated as Qt = a+bQt−1+εt,where Qt is +1 for the buy order, −1 for the sell order, and 0 for the inter-dealer orders.We write the unexpected order flow εt as Q∗

t . The second stage is estimated as ∆Pt =a + γSQ∗

t + αS∆Qt + wXt + ωt. To assess the changes of effective spreads of ETF-bondsrelative to the changes of effective spreads of non-ETF bonds, we interact ∆Qt with Postand ETF . The non-informational component αS is divided into the changes of the effectivespreads in the treatment group and the changes of the effective spreads in the control group:α1∆Qt + α2S∆Qt × Post + α3S∆Qt × ETF + α4S∆Qt × Post× ETF. Of our main focusis the coefficient α4S. We include the three public information variables Xt as discussed insection 3.5. The statistical significance is measured the probability from block bootstrapping,following Bessembinder et al. (2006).

coeff p-valueFirst stage regressionIntercept 0.1137∗∗∗ 0.000Qt−1 0.0678∗∗∗ 0.000

Second stage regressionQ∗ 0.0455∗∗∗ 0.000Treasury return −0.0138∗∗∗ 0.000Stock return × invgrd 0.0175∗∗∗ 0.000Stock return × noninvgrd 0.0740∗∗∗ 0.000Default spread × invgrd −0.0069∗∗∗ 0.000Default spread × noninvgrd 0.0383∗∗∗ 0.000∆Q 0.4160∗∗∗ 0.000∆Q× Post 0.0252∗∗∗ 0.000∆Q× ETF −0.0531∗∗∗ 0.000∆Q× Post × ETF −0.0122∗∗∗ 0.000

Firm Fixed Effect YR2 0.1880N of block bootstrapping 50

68